Interviews With Scientists

Interview with Prof. Moritz Helmstaedter: The decade-long debate has concluded, Has the European Brain Project failed?

[:en]Interview with Prof. Moritz Helmstaedter: The decade-long debate has concluded, Has the European Brain Project failed?[:zh]十年争议落幕，欧洲脑计划失败了吗？｜追问专访·Moritz Helmstaedter[:]

Interview with Professor Si Wu: Unlocking the "Door of Wisdom" in Artificial Intelligence

[:en]Interview with Professor Si Wu: Unlocking the "Door of Wisdom" in Artificial Intelligence[:zh]吴思：打开人工智能的“智慧之门”[:]

Interview with Prof. Wang Xiaojing: Fostering interdisciplinary talents for computational neuroscience

[:en]Interview with Prof. Wang Xiaojing: Fostering interdisciplinary talents for computational neuroscience[:zh] 汪小京 · 为计算神经科学培养跨学科人才[:]

Deep within the human brain, there&#8217;s a fascinating phenomenon that has captivated us for centuries—an intricate interplay between music and the brain. From soothing lullabies to exhilarating symphonies, music possesses boundless power to evoke emotions, ignite creativity, and offer us diverse sensory experiences.
&nbsp;
Scientists are currently dedicated to unraveling the mystery of how music influences human neural circuits, shedding light on its impact on various regions and functions of the brain. We are honored to have invited Professor Du Yi from Institute of Psychology, Chinese Academy of Sciences, to explore the effects of music on the brain from an anti-aging perspective.
<img loading="lazy" decoding="async" class="aligncenter wp-image-787" src="https://admin.next-question.com/wp-content/uploads/2023/11/1699193665366.png" alt="" width="600" height="274" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/1699193665366.png 717w, https://admin.next-question.com/wp-content/uploads/2023/11/1699193665366-300x137.png 300w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
<h4>Question: Please briefly introduce your research background.</h4>
&nbsp;
Professor Du: I took my undergraduate courses in medicine, but when it came to choosing my doctoral program, I developed a strong interest in psychology, especially in unraveling the mysteries of human brains. Consequently, I switched to a Ph.D. program in psychology. During my doctoral studies, my research primarily focused on auditory studies using rats as subjects, encompassing both physiological and behavioral research. However, in the later stages of my Ph.D. program and during my postdoctoral research, I became increasingly fascinated by the human brain and its processing of advanced auditory signals, such as language and music. As a result, after establishing my own laboratory, I have been dedicated to researching how the human brain comprehends language, appreciates music, and the neural mechanisms underlying these processes.
&nbsp;
As people age, the structure and function of the brain will degenerate. One significant change that impacts the quality of life for the elderly is the deterioration of hearing. In noisy environments like streets or restaurants, it becomes increasingly challenging for the elderly to discern what others are saying. This not only affects their social interactions but also their emotional well-being and even cause premature dementia. Therefore, we are deeply interested in understanding how the elderly manage speech comprehension in such noisy settings and whether they employ additional cognitive mechanisms for this purpose.
&nbsp;
In a paper we published in Nature Communications in 2016, we discovered that despite the decline in perceptual processing abilities among the elderly, they tend to engage additional brain regions, particularly those associated with speech motor control, to actively predict what others will say. This proactive, top-down prediction mechanism helps the elderly better compensate for their declining perceptual abilities, enabling them to comprehend language more effectively, especially in complex environments.
&nbsp;
Building upon this background, in our study published in PNAS in 2017, we found that young individuals, particularly young musicians, excel at speech comprehension in complex environments. We were interested in uncovering the underlying neural mechanisms, and our results revealed that young musicians possess stronger perceptual-motor integration capabilities. They rely not only on auditory input but also on predicting and compensating for speech motor movements and integrating cross-modal information to aid their speech comprehension in intricate environments.
&nbsp;
These two studies suggest a potential connection between perception and motor. It appears that the elderly may rely more on advanced, top-down perceptual-motor predictions and integration, while young musicians, attributable to their musical training experience, have advatanges in utilizing perceptual-motor integration for speech comprehension. Naturally, this leads us to inquire whether the elderly, through systematic musical training, can also enhance their cross-modal information integration abilities. Will such training potentially aid them in improving their speech comprehension within complex environments?
&nbsp;
Based on this premise, we embarked on a series of subsequent studies and found that compared to elder non-musicians, elder musicians are better at speech perception in noisy environments. In 2021, we published related behavioral research in Ear and Hearing, and followed up with new studies whose most recent findings have been published in Science Advances.
&nbsp;
<h4>Question: You mentioned that you recently published a study in Science Advances. Could you please introduce the new findings from this research?</h4>
&nbsp;
Professor Du: The recently published study in Science Advances is a featured cover article. We recruited three groups of subjects: one consisted of young non-musicians as a control group, but our primary focus was on the other two groups—elder non-musicians and elder musicians.
&nbsp;
We had them perform a visual-auditory speech syllable discrimination task in noisy environment, while monitoring their brain activity using fMRI. They would hear different syllables while simultaneously viewing matching lip movements on a screen. Their task was to determine what exactly they have heard against background noise of varying intensities.
&nbsp;
We found that behaviorally, the performance of elder musicians was closer to that of young individuals, with no significant differences in their performance, even in moderate and light noise backgrounds. In other words, musical training experience can assist the elderly in achieving performance similar to that of young individuals. This finding leads to the question: what are the underlying mechanisms?
&nbsp;
Firstly, compared to elder non-musicians, elder musicians exhibited neural processing patterns (also known as representation patterns) that were more similar to those of young individuals in both the left and right perceptual-motor regions when processing syllables in noisy environment. Sometimes, there’s even no significant differences between the two groups. In fact, their performance was significantly better than that of elder non-musicians who displayed a noticeable decline in distinguishing different speech sounds compared to young individuals. Their brains struggled to distinguish between, for example, “ba” and “da”. In contrast, elder musicians maintained strong neural differentiation abilities, similar to young individuals. This indicates that musical training experience can enhance and rejuvenate brain function which we refer to as functional preservation.
&nbsp;
Furthermore, we also discovered that elder musicians, in contrast to elder non-musicians, activated a broader range of brain regions within the frontoparietal network. These brain regions are associated with working memory and cognitive control. This suggests that elder musicians are capable of engaging more advanced brain regions for behavioral compensation.
&nbsp;
Simultaneously, elder musicians exhibited greater suppression of a brain region known as the angular gyrus compared to elder non-musicians. As a core component of the default brain network, angular gyrus is associated with autobiographical memory and long-term memory. Therefore, it tends to be activated when we are not engaged in specific tasks and relatively suppressed when we perform cognitive tasks. When we engage in cognitive tasks that require our attention to external stimuli, it becomes necessary to suppress brain regions in default network to prevent them from interfering.
&nbsp;
We found that elder musicians excel at suppressing this brain region, indicating their ability to focus more attentively on external stimuli and suppress irrelevant information. This enhanced suppression also assists them in achieving better behavioral performance. Elder musicians achieve improved performance by employing additional activation or inhibition of certain brain regions, a mechanism referred to as functional compensation.
&nbsp;
This mechanism is related to the first one mentioned earlier &#8212; functional preservation. We found that these two mechanisms are interconnected and mutually influence each other. When an elderly possess stronger abilities to activate the frontoparietal network and to inhibit the angular gyrus, his/her neural representation in the perceptual-motor brain regions also improves, rejuvenated to the level of young individuals. This suggests that functional compensation can support enhanced functional preservation. Based on these two mechanisms, musical training experience can assist elder individuals in speech perception in noisy environments and improve their behavioral performance.
<img loading="lazy" decoding="async" class="aligncenter wp-image-784" src="https://admin.next-question.com/wp-content/uploads/2023/11/81.png" alt="" width="600" height="761" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/81.png 1080w, https://admin.next-question.com/wp-content/uploads/2023/11/81-237x300.png 237w, https://admin.next-question.com/wp-content/uploads/2023/11/81-808x1024.png 808w, https://admin.next-question.com/wp-content/uploads/2023/11/81-768x974.png 768w" sizes="(max-width: 600px) 100vw, 600px" />
<h4>Question: Based on your research, how often does one need to play a musical instrument to potentially help preserve cognitive function and counteract the negative effects of aging?</h4>
&nbsp;
Professor Du: In this study, we asked the elderly to maintain a minimum of one hour of training per week within the last three years. Although they self-reported a substantial training frequency over the preceding three to four decades, it was challenging to track their training durations during the 1960s and 1970s due to historical reasons. Therefore, our focus was solely on the training frequency in the most recent three years.
&nbsp;
<h4>Question: Is it possible to achieve cognitive protection effects through temporary practice if one had no prior experience in playing a musical instrument?</h4>
&nbsp;
Professor Du: That&#8217;s a great question. What we&#8217;re currently conducting is a cross-sectional study where we directly compare the elderly with long-term musical training to those without such experience to observe differences, and we have found some. Additionally, there are numerous other longitudinal studies, both long-term and short-term, indicating that brief periods of musical training lasting from three to six months can also enhance language processing abilities for the elderly.
&nbsp;
<h4>Question: Can we achieve the same effects by listening to music?</h4>
&nbsp;
Professor Du: We all hope that passive music listening can yield similar positive effects, but current findings do not allow us to draw that conclusion definitively. Our study focuses on the effects of actively engaging in musical training, which involves multiple systems including visual, auditory, motor, emotional, and attentional processes. This multi-system engagement may have a more pronounced impact. However, listening to music also has its own benefits, such as mood regulation, stress relief, and potential improvement of focus and attention.
&nbsp;
<h4>Question: The participants in your study were engaged in a visual-auditory syllable discrimination task. Have you considered conducting other types of tasks?</h4>
&nbsp;
Professor Du: Both musicians and non-musicians in our study have completed other types of tasks. One of these tasks involved auditory working memory, where participants were presented with sequences of numbers of varying lengths and asked to repeat them in the original order or reversed order. Additionally, we assessed their executive control abilities, including the classic Stroop effect (the interference of literal meaning in font color).
&nbsp;
We found that elder musicians exhibited superior auditory working memory compared to elder non-musicians. While we didn&#8217;t observe better performance in cognitive control tasks like the Stroop effect in our study, it&#8217;s noteworthy that other research has indicated that musical experience can enhance working memory and executive control capabilities.
&nbsp;
The reason why we conducted the visual-auditory syllable discrimination task in our study was that speech processing involves real-time processing for everyone, and identifying consonants, vowels, and tones is crucial for recognition. Therefore, syllable discrimination tasks serve as a fundamental unit for speech processing.
&nbsp;
<h4>Question: Compared to non-musicians, musicians exhibit superior speech perception abilities in noisy environments. Do you believe this is solely attributable to their musical training, or could there be other factors at play?</h4>
&nbsp;
Professor Du: We chose musical training for our study because it represents a multi-system training approach that is relatively enjoyable. Moreover, it has been associated with improvements in various cognitive functions, such as working memory, selective attention, and cognitive control, among others. This includes findings from our research: stronger inhibition of the default brain network and enhanced modulation of the frontoparietal attention network among musicians. Therefore, we believe it is likely associated with musical training.
&nbsp;
However, since our study is cross-sectional, it&#8217;s important to acknowledge that there may be other differences between musicians and non-musicians in various aspects, and we cannot definitively rule out whether these differences are caused by innate factors or acquired ones.
&nbsp;
It&#8217;s possible that some individuals become musicians because of better innate structural and functional foundations or other traits that guide them toward a musical path. Therefore, our current cross-sectional study cannot definitively confirm the involvement of other factors, especially personal traits that might contribute to superior speech processing abilities. To address this question, conducting a longitudinal study would be ideal to establish whether these effects are indeed the results of musical training.
&nbsp;
<h4>Question: Are there any other studies examining similar topics? If so, are there any similarities and discrepancies between those studies and yours?</h4>
&nbsp;
Professor Du: Many researchers have indeed investigated whether musical training can assist the elderly in improving their speech perception, and they have found similar results at the behavioral level. What sets our new study apart from theirs is that we were able to identify cognitive neural mechanisms supporting both functional preservation and compensation in the brain. This allowed us to provide better explanation in terms of mechanisms to improved behavioral performance.
&nbsp;
<h4>Question: Has this study shed new light on anti-aging and brain health? Are there potential clinical application value of the findings, such as developing targeted interventions or treatment methods to enhance speech perception or other cognitive abilities among the elderly?</h4>
&nbsp;
Professor Du: We believe that musical training can have a neuroprotective effect on the brain by bringing it back to youth.
&nbsp;
While our study focused on the effects of long-term musical training, in the future, we are interested in investigating the minimum durations and frequencies of training that might also yield similar outcomes.
&nbsp;
This also reminds us that regardless of age, if you can start learning a musical instrument or even singing as early as possible, it may potentially slow down the aging process of the brain. Therefore, we currently advocate for the general public to consider learning music if plausible and to start engaging with music.
&nbsp;
In terms of clinical value, we are interested in exploring whether music therapy could be employed in the context of certain cognitive degenerative disorders among the elderly, such as Alzheimer&#8217;s disease, to improve their memory and language abilities.
&nbsp;
<h4>Question: Could music-based games potentially have similar effects?</h4>
&nbsp;
Professor Du: That&#8217;s a great approacg. Games like Rhythm Master can be effective training methods. We also refer to this as digital therapy. Using music as a form of digital therapy holds great promise for the future because gamified training like this is undoubtedly more intriguing and appealing for the elderly get fully engaged.
&nbsp;
Interviewed &amp; Edited by: Ashley Wang 
Proofread by: Xu Yunke

Interview with Prof. Du Yi: Can music rejuvenate the brain

[:en]Interview with Prof. Du Yi: Can music rejuvenate the brain[:zh]杜忆 · 音乐让大脑“返老还童”？[:]

Interview with Prof. Wu Mengyue: Do machines understand sounds better than humans?

[:en]Interview with Prof. Wu Mengyue: Do machines understand sounds better than humans?[:zh]吴梦玥 · 机器比人类更会理解声音吗？[:]

People

newsletter_sl2

newsletter_SL2

wechat_sl

wechat_SL

rss1-1

rss1 1

Follow NextQuestion

Download NextQuestion Annual report (Chinese only)

About NextQuestion

Subscribe

Contact Us

vector

Vector

x_sl

X_SL

vector-1

Vector (1)

next_question_logo_h_w_cm

Stay Up To Date With The Latest Science News

Sign Up For NextQuestion Newsletter

iconamoon_search-thin

NextQuestion

Observations

This year&#8217;s Nobel Prize curtain has drawn to a close, with the pioneers of mRNA vaccine technology clinching the 2023 Nobel Prize in Physiology or Medicine. However, for many, there remains one technology that is seen as a &#8216;missed gem&#8217; of the Nobel Prize.
Light, as a form of electromagnetic radiation, profoundly influences various aspects of life on Earth, such as photosynthesis and circadian rhythms in biological processes. This visually dazzling and colorful yet seemingly intangible light is both mysterious and captivating. In the development of neuroscience, light plays a significant role, which can be broadly categorized into three aspects. Firstly, the application of Green Fluorescent Protein (GFP), which earned the 2008 Nobel Prize in Chemistry, allows researchers to label neurons with fluorescence. Secondly, by coupling voltage-sensitive proteins on cell membranes or calcium-sensitive proteins inside cells with fluorescent proteins, researchers can monitor neuronal activity through changes in fluorescence. The third aspect is the use of light to manipulate neuronal activity, known as optogenetics.
Back in 1979, Francis Crick, who, along with James Watson, unraveled the double helix structure of DNA and was a co-recipient of the 1962 Nobel Prize in Physiology or Medicine, posited a major challenge facing neuroscience: the ability to control a specific type of cell in the brain without affecting others. Before the advent of optogenetics, physical or chemical means were primarily used to control cells. Unfortunately, existing electrodes could not precisely target specific cell groups, and the action of chemical drugs was too slow. It was not until the early 21st century that the spark of optogenetics illuminated the entire field of neuroscience. Researchers now simply need to shine light to control a tiny molecular switch, allowing them to swiftly activate or deactivate specific neurons with a &#8216;click&#8217;.
<h3>1. Optogenetics Hall of Fame</h3>
The Photoreceptive Elements Discovered in Bacteria
The story begins over half a century ago. In 1969, 29-year-old Dieter Oesterhelt, fresh from his doctoral graduation, began his postdoctoral research at the University of California, San Francisco. His project focused on a type of bacteria that thrived in high-salinity conditions, known as Halobacterium salinarum, specifically studying its purple membrane structure. During this pursuit, Oesterhelt encountered what would become his lifelong academic muse: bacteriorhodopsin. He discovered that this protein, located in the cell membranes of Halobacterium salinarum, functioned as a light-sensitive ion pump. Rapidly activated by light, this protein drove proton transport, generating chemical energy and thus enabling the bacteria to survive in low-oxygen conditions by harnessing light.
&nbsp;
<div id="attachment_1325" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1325" class="wp-image-1325" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片4-243x300.png" alt="" width="600" height="740" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片4-243x300.png 243w, https://admin.next-question.com/wp-content/uploads/2023/10/图片4.png 317w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Halobacterium salinarum NRC-1 (Scale = 270 nm). Figure Source: Wikipedia</div>
&nbsp;
Subsequently, Peter Hegemann, a student of Dieter Oesterhelt, continued the legacy of his mentor. With Dieter Oesterhelt, he isolated another type of light-sensitive ion pump, specifically the chloride ion-specific halorhodopsin, and also extended the exploration of photoreceptors to other biological systems, including Chlamydomonas reinhardtii. Building on his extensive research experience, Hegemann astutely hypothesized that the phototactic behavior of Chlamydomonas reinhardtii might be linked to a type of light-sensitive protein.
After decades of diligent research, Peter Hegemann and his team, at the beginning of the 21st century, successfully identified Channelrhodopsin-1 (ChR1) and Channelrhodopsin-2 (ChR2). Upon exposure to light, these proteins induce cell depolarization, generating a photocurrent that ultimately triggers the phototactic movement of the flagella in Chlamydomonas.
&nbsp;
<div id="attachment_1326" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1326" class="wp-image-1326" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片5-293x300.png" alt="" width="600" height="615" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片5-293x300.png 293w, https://admin.next-question.com/wp-content/uploads/2023/10/图片5.png 330w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Chlamydomonas reinhardtii. Figure Source: Wikipedia</div>
&nbsp;
The photoproteins described earlier, namely bacteriorhodopsin, halorhodopsin, and channelrhodopsins, have found applications in optogenetics. Naturally occurring bacteriorhodopsin, the first discovered member of this family that pumps protons out of the cell, and halorhodopsin, which pumps chloride ions into the cell, typically have inhibitory effects in the nervous system. This is due to the hyperpolarizing currents generated by these two proteins, which make it harder for neurons to fire action potentials. Conversely, naturally occurring channelrhodopsins, which predominantly allow positively charged ions to flow through the protein pores, tend to depolarize and excite neurons.
&nbsp;
<div id="attachment_1327" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1327" class="wp-image-1327" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片6-300x146.png" alt="" width="600" height="291" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片6-300x146.png 300w, https://admin.next-question.com/wp-content/uploads/2023/10/图片6.png 707w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Targeted activation (like activating channelrhodopsin with blue light) or inhibition (such as activating halorhodopsin with yellow light) enables rapid, cell-specific, and even projection-specific neuronal manipulation. Figure Source: Reference 1</div>
&nbsp;
Passing the Torch Among Three Forces of Power
The discovery of photoproteins may have laid an excellent foundation for the birth of optogenetics. However, the integration of these photoproteins with the manipulation of neurons still required a final, pivotal push from the trailblazers of the era. As time marched into the late 20th and early 21st centuries, three key forces played an especially vital role in propelling the emergence of optogenetics.
The first of these forces originated from the team of Austrian neuroscientist Gero Miesenböck. Miesenböck envisioned using photoproteins to manipulate neurons, but initially, he struggled to find an efficient molecular tool among them. It was not until 2002 that Miesenböck&#8217;s pioneering work demonstrated the heterologous expression of three photoreceptor genes from fruit flies in vertebrate neurons, enabling activation of specific neuron types through light stimulation.
Another pivotal group was led by Peter Hegemann and his collaborator Georg Nagel. Between 2002 and 2003, Nagel and Hegemann identified the key channelrhodopsins, ChR1 and ChR2, involved in the phototactic behavior of Chlamydomonas reinhardtii. They innovatively expressed these proteins ectopically in cells of other species, proving that their ion channel activity remained light-activated even when expressed heterologously. This method required only the expression of a single protein. The work of Nagel and Hegemann provided an incredibly simple and effective molecular tool for optogenetic techniques.
The final pivotal contribution came from the team of Edward S. Boyden and Karl Deisseroth, who masterfully integrated ChR with neuronal manipulation. In 2005, Boyden and Deisseroth reported a landmark study: by expressing just a single molecule of ChR2 in hippocampal neurons, they could trigger action potentials in specific neurons with millisecond precision through light, elegantly establishing the fundamental paradigm of optogenetics. Moreover, Boyden and Deisseroth’s team refined this tool, enabling researchers to rapidly and precisely silence specific neurons.
In 2006, Karl Deisseroth officially coined the term &#8216;optogenetics&#8217; for this tool. Optogenetics was rapidly adopted across various fields of biology due to its groundbreaking potential.
Major Awards in Optogenetics in Recent Years
In 2018, Peter Hegemann, Edward S. Boyden, and Karl Deisseroth were awarded the Canada Gairdner International Award.
In 2019, Peter Hegemann, Gero Miesenböck, Edward S. Boyden, and Karl Deisseroth received the Warren Alpert Foundation Prize.
In 2020, Peter Hegemann, Georg Nagel, and Gero Miesenböck were honored with the Shaw Prize in Life Science and Medicine.
In 2021, Dieter Oesterhelt, Peter Hegemann, and Karl Deisseroth were recipients of the Lasker Awards for Basic Medical Research.
Regrettably, Dieter Oesterhelt passed away in 2022.
&nbsp;
<div id="attachment_1328" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1328" class="wp-image-1328" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片7-300x192.png" alt="" width="600" height="384" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片7-300x192.png 300w, https://admin.next-question.com/wp-content/uploads/2023/10/图片7-768x491.png 768w, https://admin.next-question.com/wp-content/uploads/2023/10/图片7.png 865w" sizes="(max-width: 600px) 100vw, 600px" />Caption: As of 2015, a timeline of publications on photoproteins and optogenetics research. In recent years, the number of studies related to optogenetics has surged dramatically. Figure Source: Reference 12</div>
&nbsp;
<h3>2. Optogenetic Methodology Explained</h3>
Optogenetics allows for the manipulation of complex biological systems, such as freely moving mammals, with millisecond precision and cell type-specific accuracy using light. This remarkable technique involves three core components: (a) the family of photoproteins, which undergo structural changes upon light stimulation, triggering ion flow across cell membranes; (b) targeting the expression of specific rhodopsin proteins in certain cell types within specific brain regions; and (c) applying light stimulation with spatial and temporal precision to the targeted brain areas and cell types.
In a typical optogenetic experiment, the following six steps are primarily involved:
Firstly, researchers construct an expression vector containing both the ChR2 gene (or other photoproteins) and genetic elements to control its expression, such as cell type-specific promoter sequences.
Next, the expression vector is packaged into a virus.
Subsequently, this virus is injected into the animal&#8217;s brain. Although the virus infects neurons broadly, the rhodopsin proteins are expressed only in subpopulations of cells with the necessary machinery to activate the specific promoters. This results in cell type-specific expression of the light-sensitive rhodopsin proteins, which localize to the cell membranes of these neurons.
Fourthly, an optic fiber is implanted into the animal&#8217;s skull.
Fifth, light of a specific wavelength is transmitted through the optic fiber to activate these rhodopsin proteins.
Finally, researchers record and analyze the resulting electrophysiological and behavioral data.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1329" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片8-222x300.png" alt="" width="600" height="811" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片8-222x300.png 222w, https://admin.next-question.com/wp-content/uploads/2023/10/图片8.png 444w" sizes="(max-width: 600px) 100vw, 600px" /> <img loading="lazy" decoding="async" class="aligncenter wp-image-1330" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片9-192x300.png" alt="" width="600" height="937" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片9-192x300.png 192w, https://admin.next-question.com/wp-content/uploads/2023/10/图片9.png 467w" sizes="(max-width: 600px) 100vw, 600px" />
<div id="attachment_1331" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1331" class="wp-image-1331" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片10-300x223.png" alt="" width="600" height="446" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片10-300x223.png 300w, https://admin.next-question.com/wp-content/uploads/2023/10/图片10.png 339w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Basic experimental steps of optogenetics. Figure Source: Reference 15</div>
&nbsp;
Caption: Basic experimental steps of optogenetics. Figure Source: Reference 15
<h3>3. How Does Optogenetics Work?</h3>
In 2010, optogenetics was hailed as the &#8216;Method of the Year&#8217; by Nature Methods and was recognized by Science as one of the breakthroughs of the decade. The application of this technology spans two main areas: basic research on the function of neural circuits and clinical treatment of specific diseases.
Basic Research &#8211; Discovery of the &#8216;Success Breeds Success&#8217; Circuit
The first paper, published in 2007, on the application of optogenetics in animals, studied the function of neurons secreting hypocretin in the lateral hypothalamus. Researchers stimulated the activity of hypocretin-secreting neurons in the lateral hypothalamic area using optogenetics, clarifying their role in the sleep-wake state transitions in mice.
Optogenetics has also been applied to elucidate the physiological mechanisms underlying various psychiatric disorders. For instance, it has elegantly demonstrated the crucial role of cholinergic interneurons in the nucleus accumbens in cocaine-conditioned responses, the importance of midbrain dopamine neurons in chronic mild stress-induced depressive-like phenotypes, and the role of cortico-striatal neural circuits in mechanisms of compulsive behavior, as well as the involvement of the neural circuit from the medial prefrontal cortex to the basolateral amygdala in anxiety.
In 2017, Hailan Hu’s team at Zhejiang University notably used optogenetics and other techniques to validate the &#8216;winner effect&#8217; in mice. They showed that the neural circuit from the dorsomedial hypothalamus nucleus to the prefrontal cortex mediated long-term changes in social dominance status in competitive scenarios, influenced by past experiences of victory.
&nbsp;
<div id="attachment_1332" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1332" class="wp-image-1332" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片11-300x51.png" alt="" width="600" height="102" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片11-300x51.png 300w, https://admin.next-question.com/wp-content/uploads/2023/10/图片11-768x130.png 768w, https://admin.next-question.com/wp-content/uploads/2023/10/图片11.png 831w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Optogenetics Empowers the &#8216;Underdog&#8217; &#8211; Light stimulation of the thalamo-prefrontal cortex neural circuit boosts the fighting ability of mice previously disadvantaged in competitive situations. Figure Source: Reference 18</div>
&nbsp;
Clinical Treatment &#8211; &#8216;Restoring Sight&#8217; with Phototherapy
In 2021, a study published in Nature Medicine showcased, for the first time, the application of optogenetics in treating a neurodegenerative eye disease, retinitis pigmentosa, achieving partial restoration of visual function in patients. In this experiment, researchers injected a virus carrying photoproteins into the eyes of patients blinded by this disease. Subsequently, the patients&#8217; eyes were exposed to light stimulation, activating the optogenetically modified retinal ganglion cells.
Following this treatment, patients regained the ability to perceive, locate, count, and touch various objects using their vision. Multi-channel electroencephalogram recordings during this visual perception process revealed object-related activity in the visual cortex, indicating a partial recovery of visual function in these patients.
&nbsp;
<div id="attachment_1333" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1333" class="wp-image-1333" src="https://admin.next-question.com/wp-content/uploads/2023/10/图片12-300x218.png" alt="" width="600" height="436" srcset="https://admin.next-question.com/wp-content/uploads/2023/10/图片12-300x218.png 300w, https://admin.next-question.com/wp-content/uploads/2023/10/图片12-768x558.png 768w, https://admin.next-question.com/wp-content/uploads/2023/10/图片12.png 1021w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Optogenetics &#8216;Restores Sight&#8217; in the Blind. Figure Source: Reference 23</div>
&nbsp;
As highlighted in Nature Methods, using light to regulate specific activities in specific cells has achieved tremendous success in biological research, illuminating previously unexplored paths in science.

Optogenetics - A Nobel Prize Contender

[:en]Optogenetics - A Nobel Prize Contender[:zh]光遗传——诺奖的种子选手[:]

From The Journals

Professor Nan Hao from UCSD and his research group published a paper titled Engineering longevity—design of a synthetic gene oscillator to slow cellular aging in Science on April 27, 2023. The research team used yeast cells as the subjects and employed synthetic biology techniques to design a gene oscillator that successfully prolonged the life span of yeast cells by 82%.
In 2020, Professor Nan Hao&#8217;s team identified two aging patterns associated with yeast cell death, mediated by the lysine deacetylase Sir2 and heme-activated protein (HAP), respectively. At the critical point where the yeast cell declines toward death, aging is not the result of the accumulation of the two patterns, but rather the yeast cell simply “chooses” one of the aging patterns to function until it dies. About half of cellular aging is caused by a gradual decrease in the stability of rDNA, while the other half by the functional damage of mitochondria.
After further identifying the characteristics and association of the two aging patterns, the researchers began to think about whether it is possible to influence the expression of the two proteins, Sir2 and HAP4, through gene editing or chemical interventions, and thus reprogram the cellular aging pattern to slow down yeast cell aging. Professor Nan Hao&#8217;s team conducted computational simulations of the biological mechanisms of yeast cell aging, tested the experimental design, and constructed and modified gene circuits in yeast cells.
The edited yeast cells switched back and forth between high levels of Sir2 and HAP, which the researchers call a &#8220;genetic oscillator. This &#8220;genetic oscillator&#8221; drives the cells to periodically switch between two different &#8220;senescence&#8221; states, slowing down cell degradation by avoiding prolonged development of one senescence process.
The team used microfluidics coupled with time-lapse microscopy to track dynamic changes throughout the life span of single cells. And the results of the experiments showed that the yeast cells with the gene oscillator increased their life span by 82% compared to the control group of WT yeast cells, setting a new record for life span extension through genetic and chemical interventions.
Unlike previous experiments that simply knocked out or overexpressed aging-related genes, Professor Nan Hao&#8217;s team set up a gene oscillator that effectively slowed down the progression of aging in yeast cells, and significantly increased the life span of the cells. This not only marks another critical step in studying the mechanism of aging, but also lays the foundation for the use of synthetic biology to extend the life span of complex organisms.
<a href="https://www.science.org/doi/10.1126/science.add7631?url_ver=Z39.88-2003&amp;rfr_id=ori:rid:crossref.org&amp;rfr_dat=cr_pub%20%200pubmed">Read the paper on Science</a>

[:en]Professor Nan Hao from UCSD and his research group published a paper titled Engineering longevity—design of a synthetic gene oscillator to slow cellular aging in Science on April 27, 2023. The research team used yeast cells as the subjects and employed synthetic biology techniques to design a gene oscillator that successfully prolonged the life span of yeast cells by 82%.

In 2020, Professor Nan Hao's team identified two aging patterns associated with yeast cell death, mediated by the lysine deacetylase Sir2 and heme-activated protein (HAP), respectively. At the critical point where the yeast cell declines toward death, aging is not the result of the accumulation of the two patterns, but rather the yeast cell simply “chooses” one of the aging patterns to function until it dies. About half of cellular aging is caused by a gradual decrease in the stability of rDNA, while the other half by the functional damage of mitochondria.

After further identifying the characteristics and association of the two aging patterns, the researchers began to think about whether it is possible to influence the expression of the two proteins, Sir2 and HAP4, through gene editing or chemical interventions, and thus reprogram the cellular aging pattern to slow down yeast cell aging. Professor Nan Hao's team conducted computational simulations of the biological mechanisms of yeast cell aging, tested the experimental design, and constructed and modified gene circuits in yeast cells.

The edited yeast cells switched back and forth between high levels of Sir2 and HAP, which the researchers call a "genetic oscillator. This "genetic oscillator" drives the cells to periodically switch between two different "senescence" states, slowing down cell degradation by avoiding prolonged development of one senescence process.

The team used microfluidics coupled with time-lapse microscopy to track dynamic changes throughout the life span of single cells. And the results of the experiments showed that the yeast cells with the gene oscillator increased their life span by 82% compared to the control group of WT yeast cells, setting a new record for life span extension through genetic and chemical interventions.

Unlike previous experiments that simply knocked out or overexpressed aging-related genes, Professor Nan Hao's team set up a gene oscillator that effectively slowed down the progression of aging in yeast cells, and significantly increased the life span of the cells. This not only marks another critical step in studying the mechanism of aging, but also lays the foundation for the use of synthetic biology to extend the life span of complex organisms.

<a href="https://www.science.org/doi/10.1126/science.add7631?url_ver=Z39.88-2003&amp;rfr_id=ori:rid:crossref.org&amp;rfr_dat=cr_pub%20%200pubmed">Read the paper on Science</a>[:zh]“见说嵩阳有仙客，欲持金简问长生。”

长生不老，是人们从古至今的追求与期待，最直接的方法是保持健康、延缓衰老。衰老是一个目前无法终止、缓慢而复杂的过程，往往伴随着身体机能下降以及各类疾病的发生。研究衰老的生物学机制成为了众多科学家关注的重点。

美国加州大学圣地亚哥分校郝楠教授团队题为“Engineering longevity—design of a synthetic gene oscillator to slow cellular aging”的工作于2023年4月27日发表于权威期刊Science。实验人员以酵母菌作为研究对象，利用合成生物学技术设计基因振荡器成功将酵母细胞寿命延长82％。

&nbsp;

[caption id="attachment_838" align="aligncenter" width="600" class="wp-image-837 size-full"]<img class="wp-image-838" src="https://admin.next-question.com/wp-content/uploads/2023/11/212.png" alt="" width="600" height="514" /> ▷图注：Sir2和HAP介导同源酵母细胞的差异性衰老。图片来源：Yang Li, et al./ Science 2020[/caption]

&nbsp;

有趣的是，在酵母细胞功能衰退走向死亡的关键节点，衰老并非两种模式共同累计导致的结果，酵母细胞仅仅“选择”其中一条衰老通路运行直至生命消逝。即便是生活在相同环境中具有相同遗传物质的酵母细胞也会沿两种模式逐渐衰老。约一半的细胞衰老是由于rDNA的稳定性逐渐下降，而另一部分细胞则因线粒体功能损伤、衰退而衰老。

为进一步了解两种衰老模式的特点及关联，研究人员人为干预Sir2和合成HAP的主要成分HAP4在酵母细胞中的表达。特异性敲除Sir2后，HAP活性显著增加，血红素丰度上升；特异性敲除HAP4后，rDNA沉默增加，核仁构象稳定。HAP4过度表达时，大部分细胞Sir2失活，rDNA沉默减少，呈现“衰老模式1”；Sir2过度表达时，血红素含量降低，呈现“衰老模式2”的细胞比例增加，不过Sir2过表达时还出现了第三种衰老模式，即同时保持高rDNA沉默和高血红素丰度，这种细胞的寿命也更长。

&nbsp;

[caption id="attachment_839" align="aligncenter" width="600"]<img class="wp-image-839" src="https://admin.next-question.com/wp-content/uploads/2023/11/213.png" alt="" width="600" height="323" /> ▷图注：计算模型揭示了Sir2-HAP回路的多重稳定性，这使设计新的衰老模式成为可能。图片来源：Yang Li, et al./ Science 2020[/caption]

&nbsp;

这一新发现让研究人员思考是否能够通过基因编辑或化学干预等方法影响Sir2和HAP4两种蛋白表达，进而重新编程细胞衰老方式，延缓酵母细胞衰老。他们决定从基因层面改写衰老回路。为了节省时间和资源，郝楠教授团队通过计算机进行了酵母细胞衰老生物机制的模拟实验，测试实验设计，构建、修改酵母细胞中基因回路。

&nbsp;

<img class="aligncenter wp-image-840" src="https://admin.next-question.com/wp-content/uploads/2023/11/214.png" alt="" width="600" height="155" />

[caption id="attachment_841" align="aligncenter" width="600"]<img class="wp-image-841" src="https://admin.next-question.com/wp-content/uploads/2023/11/215.png" alt="" width="600" height="210" /> ▷图注：构建合成基因振荡器以重新编写酵母细胞衰老过程。图片来源：Zhen Zhou, et al./ Science 2023[/caption]

&nbsp;

为了使HAP对Sir2进行强有力的正向转录调控，团队使用细胞色素C1（CYC1）启动子取代了SIR2本地启动子，该启动子被HAP结合并激活。在删除了酵母细胞内源HAP4拷贝后，团队设计了由磷酸三酯脱氢酶启动子（TDH3）介导的HAP4基因高转录表达的构建体，同时在受Sir2调控沉默的rDNA的非转录间隔区（NTS）亦有HAP4结构整合插入。上述即为电路整合的全部内容，二者构成负反馈回路来阻止衰老过程，即HAP升高的同时激活Sir2表达，而高浓度的Sir2会反过来抑制HAP，Sir2水平下降后rDNA不再继续沉默，表达HAP4，依次循环往复。

&nbsp;

[caption id="attachment_891" align="aligncenter" width="600"]<img class="wp-image-891" src="https://admin.next-question.com/wp-content/uploads/2023/11/640-1-300x177.png" alt="" width="600" height="353" /> ▷图注：对照组细胞和合成组细胞衰老进程的动态变化对比。图片来源：Zhen Zhou, et al./ Science 2023[/caption]

&nbsp;

编辑后的酵母细胞内高水平的Sir2和HAP来回切换，像时钟的钟摆一样有节律地摆动，故而研究人员将其称为“基因振荡器”。这一“基因振荡器”驱使细胞周期性地在两种不同的“衰老”状态之间切换，避免长时间在一个衰老过程中发展下去，从而减缓细胞的退化。

&nbsp;

[caption id="attachment_843" align="aligncenter" width="600"]<img class="wp-image-843" src="https://admin.next-question.com/wp-content/uploads/2023/11/217.png" alt="" width="600" height="629" /> ▷图注：基因振荡器维持rDNA沉默和血红素生成之间的动态平衡。图片来源：Zhen Zhou, et al./ Science 2023[/caption]

&nbsp;

团队通过微流控技术和延时显微镜来追踪酵母细胞生命周期中的衰老过程，实验结果显示，与正常衰老的对照组酵母细胞相比，加设了“基因振荡器”的酵母细胞的寿命增加了82%，这一成果也创造了通过遗传和化学干预延长寿命的新纪录。

郝楠教授团队并未将衰老当做一个静态可以被彻底终结的过程，不同于以往简单敲除或过表达衰老相关基因的实验，他们设定的“基因震荡器”有效减缓了酵母细胞衰老的进展，显著增长了细胞寿命。细胞不同于电路板，由软物质构成，具有内在的变异和进化能力，即使没有外部刺激也能发生变异。

这项工作的发布，不仅标志着在衰老机制上科学家们又跨出了关键一步，同时也为使用合成生物学技术延长复杂生物寿命奠定了基础，正如Science的高级编辑L. Bryan Ray博士说的那样：“酵母细胞有一个转录切换开关，导致它们以两种命运中的一种死亡：一种是通过细胞核衰退导致死亡，另一种是通过线粒体衰亡。通过将这种转录开关重新连接成一个负反馈回路，Zhou等人能够使酵母细胞在这两种状态之间摆动，并将其寿命延长82%。这些结果代表着向使用工程原则设计控制复杂生物性状的合成基因回路迈进了一步。”

<a href="https://www.science.org/doi/10.1126/science.add7631?url_ver=Z39.88-2003&amp;rfr_id=ori:rid:crossref.org&amp;rfr_dat=cr_pub%20%200pubmed">阅读原文</a>[:]

Computational model + engineering prolongs the life span of yeast cells by 82%

[:en]Computational model + engineering prolongs the life span of yeast cells by 82%[:zh]延寿新纪录！计算机模拟+工程学，有效延长酵母细胞82%寿命[:]

[:en]
	&nbsp;



	&nbsp;



	The 21st century is celebrated as the Information Age, a time when information has become an indispensable tool for survival and interaction with the world around us. But do we truly grasp the essence of &quot;information&quot;? It drives the era forward, yet its substantial impact on our lives and thought processes remains to be fully appreciated. Claude Shannon, a distinguished mathematician, laid the groundwork for information theory, offering a theoretical framework to discuss the quantification, storage, and transmission of information.



	According to this theory, information serves as a &quot;clue&quot; that significantly reduces the uncertainty surrounding unknown events. If knowledge about certain events decreases the uncertainty of another, then that knowledge is considered information. Consider a newcomer at a company who initially knows little about it and feels uncertain about his colleagues. After sharing a few meals, he observes their behavior and gradually discerns their interests and personalities. This newfound information makes the colleagues less unfamiliar and notably diminishes the newcomer&#39;s sense of uncertainty. In our everyday lives and interpersonal interactions, information is ubiquitous, profoundly shaping our communication and cognitive processes.



	From the perspective of information theory, the human brain is seen as an information-processing entity, with information acting as the neural system&#39;s fundamental currency. Neurons exchange information through electrical and neurochemical signals, constantly refining and integrating this data within the brain to shape our cognition and memory. Thus arises the question: Are electrical and chemical signals themselves information?



	Neuroscientists have yet to agree on this matter [1]. Referring to information as the "base currency" in a rhetorical sense may not capture its true essence. Recently, information theory has embraced a new viewpoint: information is not a monolithic entity but consists of various forms. This approach, known as "information decomposition," [2] seeks to break down information into three categories: unique information, redundant information, and synergistic information. This methodology offers a fresh perspective on the nature of information, aiding in a more thorough understanding of the brain&#39;s informational architecture and cognitive processes.



	<img alt="" class="aligncenter size-full wp-image-1864" height="313" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture6.png" width="993" /> ▷ Original paper：Luppi, Andrea I., et al. &quot;Information decomposition and the informational architecture of the brain.&quot;&nbsp;Trends in Cognitive Sciences&nbsp;(2024).



	&nbsp;


<h2>
	1. Information Decomposition: Multifaceted Information
</h2>


	The three types of information&mdash;unique, redundant, and synergistic&mdash;each possess distinct characteristics and fulfill different roles in the process of information processing. Consider visual information processing as an example. The human visual system comprises central and peripheral visual fields. The peripheral visual field captures a broad range of environmental information, including the general location and blurred details of objects. The information in the peripheral visual field of each eye is unique because closing one eye (for example, the left eye) results in the loss of that eye&#39;s peripheral information, leaving the brain to rely solely on the other eye (right eye).



	For instance, while driving, peripheral vision allows the driver to notice vehicles on the sides and behind through the rearview mirror. Should the driver close his left eye due to discomfort, he might not see a vehicle approaching quickly from the left rear side, potentially leading to an unforeseen accident. In contrast, the central visual field, which is the focal area of human vision, contains most of the detailed information we perceive. For a driver, the vehicle directly ahead in the same lane is visible within the central visual field of both eyes. Closing the left eye momentarily still allows the driver, using just the right eye, to observe the vehicle ahead. Such information is categorized as &quot;redundant information,&quot; valued for its robustness since the same information is provided by multiple sources. This overrepresentation ensures that information is still accessible even if one source is compromised. However, the drawback of redundant information is its failure to fully leverage the brain&#39;s information-gathering capacity.



	Synergistic information is the final category. Stereoscopic vision cannot be produced by a single eye. How, then, does the world appear three-dimensional to us? This effect relies on the collaboration between both eyes. Due to their different positions on the head, each eye views the same object from slightly varied perspectives. The visual cortex receives two distinct two-dimensional images, and by analyzing the binocular disparity&mdash;the differences between these images&mdash;it calculates the object&#39;s distance, thereby creating a perception of depth and three-dimensionality. In driving scenarios, the cooperation of both eyes is crucial for accurately judging the distance to vehicles or obstacles ahead. Damage to either eye significantly compromises safety. Synergistic information&#39;s key feature is its efficiency, maximizing the interaction between different parts of the brain&#39;s neural system to achieve an effect where the whole is greater than the sum of its parts, thereby playing a vital role in managing complex tasks. <img alt="" class="aligncenter size-full wp-image-1863" height="431" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture7.png" width="650" /> ▷ Figure 1: The acorn and banana in the image symbolize unique information, the rectangle denotes redundant information, and the cube illustrates synergistic information, which requires the combined effort of both eyes to perceive depth. Figure source: Reference [2] As a highly complex system, the human body's approach to information decomposition not only offers a more nuanced perspective on the structure of information but also finds applications in other domains, such as cellular automata, socio-economic data analysis, and artificial neural networks [3].



	&nbsp;


<h2>
	2. How does the brain integrate information?
</h2>


	Integration is a key concept in neuroscience and cognitive science, presenting researchers with two distinct interpretations: integration as oneness and integration as cooperation. The oneness perspective views strong correlations or synchrony between activities in different brain regions, as observed in brain data analysis, as an indication of high integration. This is based on the idea that integrated elements operate as a single unit. The more synchronous and similar the activities between two elements are, the higher their level of integration is considered. Conversely, the cooperation perspective argues that when different elements complement each other, the system&#39;s information processing capacity benefits from their interactions, signifying integration. In terms of information decomposition, integration is seen as the counterpoint to unique information.



	Oneness aligns with redundant information, while cooperation aligns with synergistic information. However, traditional neuroscience often relies on correlation to deduce the integration within the nervous system, a method that struggles to distinguish between oneness and cooperation accurately. High correlation typically suggests information redundancy, whereas low correlation can indicate either uniqueness or synergistic information. To differentiate redundancy from synergy, researchers have developed the synergy-redundancy index, which reflects the balance between these two in a system. If the system&#39;s total information output exceeds the sum of its individual parts, synergy is inferred. If the sum of the parts&#39; contributions exceeds the system&#39;s total output, redundancy is present. While intuitive, this approach has limitations in situations where synergy and redundancy coexist, and it cannot precisely identify synergistic information.



	Unlike correlation-based methods or the synergy-redundancy index, information decomposition provides a more accurate analysis of information processing by calculating the system&#39;s transfer entropy (TE). For instance, if TE data show a higher transfer from time series data X to Y than from Y to X in a system, X is deemed to have a significant &quot;influence&quot; on Y. Considering epilepsy, characterized by disordered functional connectivity in the brain, traditional interpretations view the highly synchronous signals among patients as indicative of oneness. This suggests that redundancy in different brain regions could contribute to epileptic seizures. However, when analyzing EEG data from epilepsy patients using information decomposition, researchers discovered increases in both redundant and synergistic information transfer from subcortical to cortical regions during seizures. Furthermore, an analysis of deep electrode recordings revealed that an increase in unique information transfer from specific subcortical areas to cortical regions might be a primary trigger for cortical oscillations, offering clearer evidence for identifying seizure onset zones [4]. <img alt="" class="aligncenter size-full wp-image-1862" height="753" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture8.png" width="954" /> ▷ Figure 2: Information decomposition offers a unified framework for cognitive science, where each bi-directional arrow crossing the central triangular area represents opposing concepts in cognitive and neuroscience. One end of the arrow corresponds to a category within the information decomposition framework, and the other end differentiates between two types of information, such as integration and disintegration&mdash;the latter implying unique information, and the former encompassing both redundant and synergistic information. Localization refers to instances where an information source solely contains unique information, making it exclusively available from that source. In contrast, sources with only redundant information disseminate their information across multiple locations. Differentiation occurs when system parts exhibit distinct behaviors or do not demonstrate oneness, indicating they might be independent or complementary. Figure source: Reference [2]



	&nbsp;


<h2>
	3. How does the brain balance redundancy and synergy?
</h2>


	Redundancy and synergy, significant modes of interaction within the brain, have been distinguished by extensive research. A meta-analysis utilizing the NeuroSynth database, which includes over 15,000 imaging studies, revealed that redundant information is crucial for sensory-motor processing. As the brain&#39;s input-output mechanism, stable sensory-motor processing is essential for survival, with redundant interactions providing the necessary robustness. Conversely, synergistic information serves as a &quot;global workspace&quot; in the brain, crucial for executing higher cognitive functions. Areas of the brain exhibiting high synergy demonstrate increased aerobic glycolysis rates and a broader range of neurotransmitter receptor expressions, supporting flexible and rapid energy supply, synapse formation, and neural regulation [5]. At the macroscopic scale, human resting-state fMRI data indicate that synergy is more prevalent overall [5], although specific brain regions vary. The frontal and parietal association cortices, integral for integrating multimodal information, primarily operate through synergy. The frontal association cortex is involved in long-term planning and decision-making, while the parietal association cortex focuses on spatial localization, hand-eye coordination, and other tasks. In contrast, redundancy is more common in primary cortical areas that process unimodal information, such as the visual, somatomotor, and auditory cortices.



	Compared to other primates, humans possess a more developed association cortex, suggesting a greater integration of information through cooperation and synergy, underscoring our cognitive advantages. Microscale research corroborates these macroscopic findings. Electrophysiological recordings in the prefrontal cortex reveal complex and flexible neuronal responses to stimuli and task changes, highlighting the importance of neuronal interactions. However, in the visual cortex, especially in areas V4 and V1, neuronal interactions contribute less to activity patterns [6, 7]. Artificial neural network research further supports these observations, showing that redundancy is predominant in early network stages. As learning progresses, neurons specialize, contributing more unique information. When networks learn multiple tasks, requiring flexibility to integrate different information sources, synergy strengthens.



	Damage to highly synergistic neurons significantly impairs network performance. Conversely, training with randomly deactivated neurons enhances redundancy and robustness, improving resistance to damage [8]. In summary, while redundancy offers robustness and underpins sensory-motor functions in humans and many primates, it does not maximize the neural system&#39;s information-processing capacity. Synergy, associated with higher-order information processing, is more efficient and flexible, playing a crucial role in advancing human cognitive capabilities. Damage to parts of this cooperative network is likely to impair higher cognitive functions. <img alt="" class="aligncenter size-full wp-image-1861" height="910" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture9.png" width="835" /> ▷ Figure 3: (A) The image shows brain areas where redundancy (blue) and synergy (red) predominate. Blue areas, related to primary sensory-motor processing, exhibit a modular network organization with structured connections for single modality processing, like visual information in the visual cortex. Red areas, responsible for complex cognitive processing, are associated with synaptic density, dendritic spine-related genes, and enhanced plasticity. (B) The evolutionary trajectory of synergy. While redundant information remains relatively stable across species, the human brain shows more developed synergistic information, likely due to expanded cortical areas, compared to other primates. Figure source: Reference [2]



	&nbsp;


<h2>
	4. A New Blueprint for Artificial Intelligence Design
</h2>


	The framework of information decomposition not only provides a comprehensive lens for examining brain information processing but also sheds light on human evolutionary advantages. Moreover, it paves the way for researchers to design AI systems that more closely mimic human intelligence. Artificial neural networks have shown impressive capabilities across various domains. A pivotal question emerges: Do these systems rely on synergy, akin to the human brain? Recent advances in artificial intelligence have largely been driven by increases in model scale. Researchers note that as AI models grow in scale and their ability to handle multitasking improves, they display more synergistic behaviors, hinting at AI systems becoming more human-like. However, the inherent fragility of synergistic actions introduces risks for AI, suggesting that future designs of AI systems should differentiate between types of information, and evolve the information decomposition framework into a universal tool for deciphering complex systems. This approach could also illuminate the inner workings of AI models, often considered &quot;black boxes.&quot; Conversely, artificial intelligence offers a fertile ground for testing theories of information. Observations indicate that synergy intensifies when confronting complex challenges. If AI systems were engineered to favor synergistic actions, perhaps through evolutionary algorithms, might they navigate complex tasks with greater efficacy? Moreover, an AI system designed solely for synergistic actions would provide a unique lens through which to examine the benefits and limitations of synergy&mdash;a scenario beyond the reach of any biological system.



	<img alt="" class="aligncenter size-large wp-image-1860" height="652" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture10-1024x652.png" width="1024" /> ▷ Figure 4: Envisioning the information decomposition framework as a Rosetta Stone bridging biology and artificial intelligence. Just as the Rosetta Stone, inscribed with Greek, Ancient Egyptian, and other scripts, was vital for decoding Ancient Egyptian history, information processing and decomposition elucidate the connections between brain structure, functional organization, and cognitive and behavioral outcomes. Similarly, in AI systems, this framework can clarify the links between system architecture, computational abilities, and performance outcomes. Whether in the biological brain or an AI system, information processing and decomposition, transcending the medium, can serve as a universal language. Figure source: Reference [2]



	&nbsp;


<h2>
	5. Conclusion
</h2>


	The rapid advancements in artificial intelligence technology often astonish us in our everyday lives. However, the complexity of the human brain vastly exceeds that of current AI capabilities, and the framework of information decomposition brings us one step closer to deciphering the brain&#39;s enigmas. What techniques will we uncover in the future to further decompose information? How can we leverage information decomposition to develop more sophisticated AI systems? Could a deeper insight into the brain&#39;s information architecture enable us to address the myriad mental health issues affecting society? Perhaps, with enough understanding of the brain&#39;s fundamental mechanisms, the creation of an artificial brain might transition from fantasy to reality.



	&nbsp;



	&nbsp;

[:zh]

[:]

When we acquire information, what exactly do we gain?

[:en]When we acquire information, what exactly do we gain?[:]

&nbsp;


	&nbsp;


	In the early stages of human societal development, communities were primarily close-knit, where every word and action of an individual spread among neighbors, creating tightly connected small communities. In these settings, interactions among individuals were not accidental but characterized by long-term continuity. Everyone consistently acted in ways that would shape their image and reputation within the community. This continuous interaction, resembling an expansive network, tightly interwove each individual&rsquo;s fate with the others&#39;. Against this backdrop, cooperation and mutual assistance emerged as crucial cornerstones in the early development of human societies.


	This perspective is rooted in the theory of the evolution of cooperation through repeated interactions, which posits that acts of cooperation are based on the anticipation of future cooperative opportunities. Such long-term social engagements foster cooperation among individuals. However, this seemingly plausible explanation overlooks a critical detail: even within long-term interactions, individuals may engage in ambiguous reciprocal behaviors (i.e., the fuzzy reciprocation strategy), potentially undermining the foundation of cooperation. In other words, the mere expectation of future interactions does not ensure the stability of cooperation.


	Another hypothesis, group competition, suggests that competition among groups fosters the evolution of cooperation. Yet, this theory also has its limitations. While inter-group competition can encourage internal cooperation, its effectiveness wanes when groups become exceedingly similar. Furthermore, even with discernible differences between groups, group competition may not efficiently bolster cooperation since cooperative groups tend to compete with each other rather than with non-cooperative groups.


	A recent article in Nature addresses the limitations of these theories, illustrating that their integration can produce a synergistic effect. While the basis for cooperation in repeated interactions is fragile, group competition can mitigate the negative impact of ambiguous reciprocation strategies. Indeed, evolutionary strategies often involve cooperative behaviors towards in-group members and non-cooperative behaviors towards out-group members.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1847" height="473" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1.png" width="967" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1.png 967w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1-300x147.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1-768x376.png 768w" sizes="(max-width: 967px) 100vw, 967px" /> 
	▷ Paper: Efferson, Charles, et al. &quot;Super-additive cooperation.&quot;&nbsp;Nature&nbsp;(2024): 1-8.


	The experiment was conducted in West Papua, New Guinea, where the research team selected two local communities, the Ngenika and the Perepka, to participate in a two-person sequential social dilemma game. Unlike the traditional prisoner&#39;s dilemma game, this experiment&#39;s game allowed for cooperative actions to vary continuously within a certain range, more accurately reflecting the social dilemmas humans encounter in the real world.


	In this game, each participant received a certain amount of &quot;resources.&quot; The game unfolded in two steps: initially, the first mover could choose to transfer some of their resources to the second mover, which would double during the transfer; then, based on the initial transfer, the second mover could choose to transfer some of their resources back to the first mover, which would also double. This setup not only presented an opportunity for cooperation but also posed a strategic challenge.


	The experiment explored three scenarios: one featuring only repeated interaction, one with only group competition, and one that combined both. Additionally, the study examined variations in strategic space dimensions to simulate potential cooperative strategies.


	The results revealed that participants engaged in higher initial resource transfers (indicating more cooperative behavior) when paired with in-group members, and lower initial resource transfers (indicating less cooperative behavior) when paired with out-group members. Importantly, the response strategies of the second movers varied significantly: responses to in-group members often increased cooperation (by augmenting resource transfers to reward the first mover&rsquo;s cooperation), while responses to out-group members generally reduced cooperation (by decreasing resource transfers in reaction to the first mover&rsquo;s level of cooperation).


	This pattern aligned perfectly with the research team&rsquo;s model predictions, suggesting that the combination of repeated interaction and group competition could create a powerful synergistic effect, fostering in-group cooperation and mitigating the negative impact of ambiguous reciprocity on cooperation.


	This experiment challenges the notion that a single mechanism, such as repeated interactions or group competition alone, can account for human cooperation. It introduces a new perspective: human cooperation may have evolved under the combined influences of repeated interaction and group competition. The findings demonstrate that even in one-off interactions, cooperative behaviors are shaped by the historical legacy effects of these two mechanisms, implying that the cooperative behaviors we observe today may be a direct result of social motivations that evolved in the distant past through in-group repeated interactions and inter-group competition.


	From the Ngenika to the Perepka, from forests to cities, and from ancient tribes to modern societies, the essence of cooperation has consistently permeated human interaction. This experiment serves as a mirror, reflecting the intricate and nuanced evolutionary backdrop of human cooperative behavior. The essence of human cooperation is not a product of simple algorithms but is cultivated through the complexities of repeated interactions and honed in the crucible of group competition.


	&nbsp;


	&nbsp;

[:en]
	&nbsp;



	&nbsp;



	In the early stages of human societal development, communities were primarily close-knit, where every word and action of an individual spread among neighbors, creating tightly connected small communities. In these settings, interactions among individuals were not accidental but characterized by long-term continuity. Everyone consistently acted in ways that would shape their image and reputation within the community. This continuous interaction, resembling an expansive network, tightly interwove each individual&rsquo;s fate with the others&#39;. Against this backdrop, cooperation and mutual assistance emerged as crucial cornerstones in the early development of human societies.



	This perspective is rooted in the theory of the evolution of cooperation through repeated interactions, which posits that acts of cooperation are based on the anticipation of future cooperative opportunities. Such long-term social engagements foster cooperation among individuals. However, this seemingly plausible explanation overlooks a critical detail: even within long-term interactions, individuals may engage in ambiguous reciprocal behaviors (i.e., the fuzzy reciprocation strategy), potentially undermining the foundation of cooperation. In other words, the mere expectation of future interactions does not ensure the stability of cooperation.



	Another hypothesis, group competition, suggests that competition among groups fosters the evolution of cooperation. Yet, this theory also has its limitations. While inter-group competition can encourage internal cooperation, its effectiveness wanes when groups become exceedingly similar. Furthermore, even with discernible differences between groups, group competition may not efficiently bolster cooperation since cooperative groups tend to compete with each other rather than with non-cooperative groups.



	A recent article in Nature addresses the limitations of these theories, illustrating that their integration can produce a synergistic effect. While the basis for cooperation in repeated interactions is fragile, group competition can mitigate the negative impact of ambiguous reciprocation strategies. Indeed, evolutionary strategies often involve cooperative behaviors towards in-group members and non-cooperative behaviors towards out-group members.



	<img alt="" class="aligncenter size-full wp-image-1847" height="473" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1.png" width="967" /> 
	▷ Paper: Efferson, Charles, et al. &quot;Super-additive cooperation.&quot;&nbsp;Nature&nbsp;(2024): 1-8.



	The experiment was conducted in West Papua, New Guinea, where the research team selected two local communities, the Ngenika and the Perepka, to participate in a two-person sequential social dilemma game. Unlike the traditional prisoner&#39;s dilemma game, this experiment&#39;s game allowed for cooperative actions to vary continuously within a certain range, more accurately reflecting the social dilemmas humans encounter in the real world.



	In this game, each participant received a certain amount of &quot;resources.&quot; The game unfolded in two steps: initially, the first mover could choose to transfer some of their resources to the second mover, which would double during the transfer; then, based on the initial transfer, the second mover could choose to transfer some of their resources back to the first mover, which would also double. This setup not only presented an opportunity for cooperation but also posed a strategic challenge.



	The experiment explored three scenarios: one featuring only repeated interaction, one with only group competition, and one that combined both. Additionally, the study examined variations in strategic space dimensions to simulate potential cooperative strategies.



	The results revealed that participants engaged in higher initial resource transfers (indicating more cooperative behavior) when paired with in-group members, and lower initial resource transfers (indicating less cooperative behavior) when paired with out-group members. Importantly, the response strategies of the second movers varied significantly: responses to in-group members often increased cooperation (by augmenting resource transfers to reward the first mover&rsquo;s cooperation), while responses to out-group members generally reduced cooperation (by decreasing resource transfers in reaction to the first mover&rsquo;s level of cooperation).



	This pattern aligned perfectly with the research team&rsquo;s model predictions, suggesting that the combination of repeated interaction and group competition could create a powerful synergistic effect, fostering in-group cooperation and mitigating the negative impact of ambiguous reciprocity on cooperation.



	This experiment challenges the notion that a single mechanism, such as repeated interactions or group competition alone, can account for human cooperation. It introduces a new perspective: human cooperation may have evolved under the combined influences of repeated interaction and group competition. The findings demonstrate that even in one-off interactions, cooperative behaviors are shaped by the historical legacy effects of these two mechanisms, implying that the cooperative behaviors we observe today may be a direct result of social motivations that evolved in the distant past through in-group repeated interactions and inter-group competition.



	From the Ngenika to the Perepka, from forests to cities, and from ancient tribes to modern societies, the essence of cooperation has consistently permeated human interaction. This experiment serves as a mirror, reflecting the intricate and nuanced evolutionary backdrop of human cooperative behavior. The essence of human cooperation is not a product of simple algorithms but is cultivated through the complexities of repeated interactions and honed in the crucible of group competition.



	&nbsp;



	&nbsp;

[:zh]


[:]

Nature: United by a Common Foe or Sharing Scarce Resources—What Drove the Evolution of Cooperation

[:en]Nature: United by a Common Foe or Sharing Scarce Resources—What Drove the Evolution of Cooperation[:]

How Does the Large Language Model Proclaim the Death of Psychology?

[:en]How Does the Large Language Model Proclaim the Death of Psychology?[:zh]万字长文，大语言模型如何宣告心理学的死亡?[:]

<h3>The “motor homunculus” theory of the brain</h3>
&nbsp;
On the surface of the brain, the first somatic motor area is a cortical region located at the posterior convexity of the precentral gyrus (PCG), which projects innervations to the motor activities of the lower limbs, upper limbs, and head and face in an internal-to-external, top-to-bottom order, respectively. In 1937, Professor William Penfield coined the term “motor homunculus” to describe the projective innervation of the cerebral cortex.
&nbsp;
According to Gerwin Schalk, the Director of the Chen Frontier Lab for Applied Neurotechnology and also a well-known scientist in the field of motor control and brain-computer interface, “The evolving neuroscience shows us clearly this intricate map of the cerebral cortex, where at the precentral gyrus, different areas correspond to different parts of the body and project innervations to the autonomous activity of that area.”
&nbsp;
However, the “motor homunculus” theory did not further explain the coordination mechanism of movement. Later, scientists including Pandya proposed the concept of motor association area in 1982, which suggests that the motor association area is not directly involved in projecting and dominating somatic movements but do so by coordinating and connecting the motor cortex areas to achieve more complex motor regulation.
&nbsp;
<h3>New discovery: a motor association area in the depths of the central sulcus</h3>
&nbsp;
To deepen understanding of the structure and functions of the precentral gyrus, Dr. Gerwin Schalk conducted a research based on sEEG (stereoelectroencephalography) in collaboration with Michael A. Jensen along with other experts. A paper describing their findings was published in the famous journal Nature Neuroscience.
&nbsp;
The study found that there is another new region in the brain called Rolandic motor association area (RMA) which also controls and coordinates our motor functions!
&nbsp;
A total of 13 patients with drug-resistant partial epilepsy aged between 11-20 years were recruited for this study, each of whom underwent placement of sEEG electrode leads for seizure network characterization. Subjects were visually cued to perform simple self-paced (~1 Hz) movements in response to images randomly selected during which time the researchers would record signals of neural electrical activity in the brain tissues surrounding the electrodes. EMG was measured from the forearm flexors/extensors (hand), base of chin (tongue) and anterior tibialis (foot) during the motor task to identify the intervals between the cortical firing and somatic movements.
&nbsp;
A simple analysis of broadband changes extended the classic somatotopic representation of individual body parts into the sulcal depths of the PCG. Unexpectedly, a new region in the depths of the central sulcus, at its midlateral aspect, was electrophysiologically active during all three movement types. The researchers have thus named it “Rolandic motor association area (RMA)”.
&nbsp;
Dr. Schalk introduced that the organized topology of the homunculus was interrupted by this specific region. They have also discovered that the RMA areas lacks the somatotopically specific organization found in the homunculus. Because the RMA is not plainly related to any single movement function, it is likely an association area that helps to coordinate different effectors of movement.
&nbsp;
<h3>What’s significance of this discovery?</h3>
&nbsp;
Dr. Schalk said: “On this basis, the discovery of the RMA has once again expanded our knowledge of the human brain. Future studies may further explore the interconnections between the motor association area and the first somatic motor area, and the intercorrelations amongst the motor association area. This may contribute to new discoveries of RMA’s role in neural circuits.”
&nbsp;
Cutting-edge neurotechnologies were of tremendous help in identifying RMA. “With the latest advances in sensor technology, increased computing power, and the growing sophistication of signal processing/AI algorithms, we now have powerful tools to learn more about and modify brain function,” said Dr. Schalk.
&nbsp;
<a href="https://www.nature.com/articles/s41593-023-01346-z">Read the paper on Nature Neuroscience</a>

[:en]<h3>The “motor homunculus” theory of the brain</h3>
&nbsp;

On the surface of the brain, the first somatic motor area is a cortical region located at the posterior convexity of the precentral gyrus (PCG), which projects innervations to the motor activities of the lower limbs, upper limbs, and head and face in an internal-to-external, top-to-bottom order, respectively. In 1937, Professor William Penfield coined the term “motor homunculus” to describe the projective innervation of the cerebral cortex.

&nbsp;

According to Gerwin Schalk, the Director of the Chen Frontier Lab for Applied Neurotechnology and also a well-known scientist in the field of motor control and brain-computer interface, “The evolving neuroscience shows us clearly this intricate map of the cerebral cortex, where at the precentral gyrus, different areas correspond to different parts of the body and project innervations to the autonomous activity of that area.”

&nbsp;

However, the “motor homunculus” theory did not further explain the coordination mechanism of movement. Later, scientists including Pandya proposed the concept of motor association area in 1982, which suggests that the motor association area is not directly involved in projecting and dominating somatic movements but do so by coordinating and connecting the motor cortex areas to achieve more complex motor regulation.

&nbsp;
<h3>New discovery: a motor association area in the depths of the central sulcus</h3>
&nbsp;

To deepen understanding of the structure and functions of the precentral gyrus, Dr. Gerwin Schalk conducted a research based on sEEG (stereoelectroencephalography) in collaboration with Michael A. Jensen along with other experts. A paper describing their findings was published in the famous journal Nature Neuroscience.

&nbsp;

The study found that there is another new region in the brain called Rolandic motor association area (RMA) which also controls and coordinates our motor functions!

&nbsp;

A total of 13 patients with drug-resistant partial epilepsy aged between 11-20 years were recruited for this study, each of whom underwent placement of sEEG electrode leads for seizure network characterization. Subjects were visually cued to perform simple self-paced (~1 Hz) movements in response to images randomly selected during which time the researchers would record signals of neural electrical activity in the brain tissues surrounding the electrodes. EMG was measured from the forearm flexors/extensors (hand), base of chin (tongue) and anterior tibialis (foot) during the motor task to identify the intervals between the cortical firing and somatic movements.

&nbsp;

A simple analysis of broadband changes extended the classic somatotopic representation of individual body parts into the sulcal depths of the PCG. Unexpectedly, a new region in the depths of the central sulcus, at its midlateral aspect, was electrophysiologically active during all three movement types. The researchers have thus named it “Rolandic motor association area (RMA)”.

&nbsp;

Dr. Schalk introduced that the organized topology of the homunculus was interrupted by this specific region. They have also discovered that the RMA areas lacks the somatotopically specific organization found in the homunculus. Because the RMA is not plainly related to any single movement function, it is likely an association area that helps to coordinate different effectors of movement.

&nbsp;
<h3>What’s significance of this discovery?</h3>
&nbsp;

Dr. Schalk said: “On this basis, the discovery of the RMA has once again expanded our knowledge of the human brain. Future studies may further explore the interconnections between the motor association area and the first somatic motor area, and the intercorrelations amongst the motor association area. This may contribute to new discoveries of RMA’s role in neural circuits.”

&nbsp;

Cutting-edge neurotechnologies were of tremendous help in identifying RMA. “With the latest advances in sensor technology, increased computing power, and the growing sophistication of signal processing/AI algorithms, we now have powerful tools to learn more about and modify brain function,” said Dr. Schalk.

&nbsp;

<a href="https://www.nature.com/articles/s41593-023-01346-z">Read the paper on Nature Neuroscience</a>[:zh]<h3>大脑的“运动侏儒”理论</h3>
&nbsp;

约一个世纪前，神经外科医生通过对开颅患者直接进行脑表面的电刺激，首次发现了躯体运动代表区。在大脑表面，第一躯体运动区是位于中央前回（Precentral gyrus，PCG）后面凸的大脑皮质区域，按照从内到外、从上到下的顺序分别投射支配下肢、上肢和头面部的运动。1937年，William Penfield教授用经典的“运动侏儒（Motor homunculus）”模式图来描述大脑皮层的这种投射支配模式[1]。后来，“运动侏儒”模式也陆续得到了多项基于功能磁共振成像（fMRI）、脑磁图（MEG）、脑皮层电图（ECoG）等方法进行的研究所证实。

&nbsp;

Gerwin Schalk博士是运动控制神经基础与脑-机接口（BCI）领域的知名科学家，也是天桥脑科学研究院（Tianqiao and Chrissy Chen Institute，TCCI）应用神经技术前沿实验室主任。他表示：“不断发展的神经科学向我们清晰地展示了大脑皮层这张复杂而精细的地图，在中央前回处，不同区域对应躯体的不同部位，并投射支配该区域的自主活动。”

&nbsp;

[caption id="attachment_828" align="aligncenter" width="600"]<img class="wp-image-828" src="https://admin.next-question.com/wp-content/uploads/2023/11/0001-300x152.png" alt="" width="600" height="304" /> ▷图注：“感觉侏儒”（左）与“运动侏儒”（右） 图片来源：https://www.ebmconsult.com/[/caption]

&nbsp;

但“运动侏儒”理论仅仅解释了运动的产生，却没有进一步解释运动的协调机制。鉴于此，Pandya等[2]于1982年提出了运动联合区域（motor association area）概念，认为运动联合区域并不直接参与投射支配躯体运动，而是通过协调联系各运动皮层区，从而实现更加复杂的运动调控。

&nbsp;
<h3>发现新脑区：中央沟深部的运动联合区域</h3>
&nbsp;

为了更深刻地理解中央前回的结构与功能，Gerwin Schalk博士与梅奥诊所（Mayo Clinic）的学者Michael A. Jensen等众多专家共同开展了一项基于立体脑电图（stereoelectroencephalography，sEEG）的研究。该研究成果于2023年5月发表在著名期刊Nature Neuroscience上。

&nbsp;

研究发现，大脑的皮层地图并不完全像先前大家所描述的那样，大脑中其实还存在着另一个全新的脑区——罗朗多氏运动联合区域（Rolandic motor association area，RMA），这个区域同样控制和协调着我们的运动功能！

&nbsp;

研究一出，在科学界内掀起了不小的轰动。

&nbsp;

[caption id="attachment_829" align="aligncenter" width="600"]<img class="wp-image-829" src="https://admin.next-question.com/wp-content/uploads/2023/11/0002.png" alt="" width="600" height="289" /> ▷文章报道封面 图片来源：https://www.nature.com/articles/s41593-023-01346-z[/caption]

&nbsp;

本次研究共招募了13名年龄在11-20岁之间的药物难治性局灶性癫痫患者（其中有6名女性）。研究人员向患者脑内植入立体脑电图深部电极，随后指导患者根据随机抽取的图案提示，以约1次/秒的节律分别完成以下简单的自主运动：单手握紧/松开、闭嘴时舌头左移/右移、单足跖屈/背伸。在此期间，研究者会采集电极周围脑组织的神经电活动信号，计算其功率谱密度（power spectral density），并进行时间进程分析（time course analysis）。研究人员同时在前臂伸/屈肌、下颌底部、胫前肌处放置肌电电极，以评价从大脑皮层运动区放电到肢体活动间的间隔时间。研究原计划是同时记录脑表面及深部的电信号，并采用K均值聚类算法（K-means clustering algorithm）计算手、舌、脚分别活动时被激活的区域/簇在脑中的相对空间位置，从而从三维容积角度阐释支配运动过程中第一躯体运动区的脑电活动特征。

&nbsp;

[caption id="attachment_1216" align="aligncenter" width="600"]<img class="wp-image-1216" src="https://admin.next-question.com/wp-content/uploads/2023/11/图片-2-300x168.png" alt="" width="600" height="335" /> ▷图c展示了利用K均值聚类算法计算患者活动手、舌、脚时所采集脑电数据的结果。所有数据分为5个簇,以不同颜色标记并在三维坐标系中展示。图d则展示了各簇在脑内的空间位置分布。图片来源：Jensen, et al./ Nature Neuroscience 2023[/caption]

在对受试者宽频脑电图结果进行分析后发现，第一躯体运动区实际上比经典“运动侏儒”模式图描述的范围要更大，其中，部分运动代表区延伸入中央沟内。更令人惊讶的是，在所有患者按照指令活动手、舌、脚的过程中，立体脑电图显示，位于中央沟深部中外侧部的一处全新的脑区均被激活，这与周围特化的躯体运动代表区的电生理活动模式（特定脑区激活对应特定躯体部位自主活动）有着显著区别。

&nbsp;

研究者借鉴“罗朗多氏裂（Fissure of Rolando）”这一中央沟的旧称，将这个脑区命名为“罗朗多氏运动联合区域（Rolandic motor association area，RMA）”。

&nbsp;

据Schalk博士介绍，这个区域的出现打断了中央前回“运动侏儒”地图的连续性。他们所观察到患者在完成不同指令过程时RMA均被激活，这说明RMA支配运动不具有特异性，其并不通过神经纤维直接投射支配某一部位的运动，而是参与运动的协调。

&nbsp;
<h3>发现新的脑区，有什么意义？</h3>
&nbsp;

随着立体脑电图技术在临床上的普及，研究者们不断发现全新的、参与调节运动的运动联合区域。Glasser等[4]发现，位于RMA前上方的中央前回55b区参与生成语言与协调音律。而Willett等[5]指出，四肢瘫患者的运动前区（BA6区）参与整合全身的自主运动。Gordon等[6]则进一步提出，中央前回上存在三个未直接参与投射支配躯体运动的区域，这些区域其中之一与上述中央前回55b区相重叠，可能与纹状体、中央中核相联系，参与协调全身的自主运动。

&nbsp;

谈到此次发现的意义，Schalk博士激动地表示：“在这基础上，发现RMA再次扩展了我们对人脑的认知。未来研究或可进一步探讨运动联合区域与第一躯体运动区、运动联合区域之间的相互联系，从而发现运动联合区域在神经环路中的更广泛作用。“

&nbsp;

在发现RMA的过程中，先进的神经技术功不可没。作为一种日趋成熟的技术，立体脑电图能更好地从三维角度描述癫痫患者脑网络特征，有助于指导癫痫治疗。相较而言，传统的脑皮层电图与直接电刺激则很容易忽略RMA，这是因为，这些技术主要记录大脑浅表的神经电活动，而RMA位于中央沟深部，为BA4区所掩盖，其电活动相对不易被捕捉到。

&nbsp;

正是立体脑电图的立体定位、能捕捉深部白质神经电活动的优势促成了新脑区的发现。在技术层面，Gerwin Schalk博士也对先进神经技术的应用充满期待，他表示“随着传感器技术的最新进步、计算能力的增强以及信号处理/AI算法的日益复杂，我们现在拥有强大的工具来更多地了解和修改大脑功能。”

&nbsp;

目前，人们越来越深刻认识到神经技术的潜力，并开始通过系统的工作将神经科学发现与技术相结合，以开发使医疗与科研行业内外更多人受益的解决方案。“我们与梅奥诊所的这项研究有助于更好地利用这些工具来解决不同神经系统疾病的破坏性影响。例如，我们目前的发现可能会促成新的或改进的方法来改善不同运动障碍的治疗，如帕金森病或图雷特综合征。”他在采访中表示。

&nbsp;

<a href="https://www.nature.com/articles/s41593-023-01346-z">阅读原文</a>[:]

The discovery of a new motor association area reshapes human understanding of the brain!

[:en]The discovery of a new motor association area reshapes human understanding of the brain![:zh]发现人脑运动协调新区域，人类认知再次被刷新！[:]

Features

Karl Friston: Building the Future's Intelligent Ecosystem Based on Free Energy Theory

[:en]Karl Friston: Building the Future's Intelligent Ecosystem Based on Free Energy Theory[:]

Research has shown that people prefer Chatbots with certain social traits, and the question that arises is: what social traits improve a Chatbot&#8217;s communication and social skills? In what areas do these features have significant benefits?
Three years before AI topics trending on every media platforms, researchers including Chaves and Gerosa from Northern Arizona University and Federal University of Paraná had conducted an in-depth literature review. It turned out that there are 11 social characteristics in specific domains facilitating better interaction between chatbots and users, but there are still some usage limitations and challenges as well.
What social characteristics should chatbots be equipped with?
The researchers proposed 11 social characteristics necessary for chatbots and categoried them into three types, namely, conversational intelligence, social intelligence and personification.
Conversational intelligence
Conversational intelligence enables the chatbot to actively participate in the conversation and to demonstrate awareness of the topic discussed, the evolving conversational context, and the flow of the dialogue. Social characteristics related to conversational intelligence include:
<ol>
<li>Proactivity: autonomously act on the user&#8217;s behalf;</li>
<li>Conscientiousness: attentiveness to the conversation at hand;</li>
<li>Communicability: the capacity of a software to convey to users its underlying design intent and interactive principles.</li>
</ol>
Social intelligence
Social Intelligence refers to the ability of an individual to produce adequate social 
behavior for the purpose of achieving desired goals. When developing chatbots, it is necessary to account for the socially acceptable protocols for conversational interactions. Social characteristics related to social intelligence include:
<ol>
<li>Damage control: the ability to deal with either conflict or failure situations;</li>
<li>Thoroughness: to be precise regarding how it uses language;</li>
<li>Manners: to manifest polite behavior and conversational habits;</li>
<li>Moral agency: the ability to act based on social notins of right and wrong;</li>
<li>Emotional intelligence: the ability to appraise and express feelings, regulate affective reactions, and harness emotions to solve a problem;</li>
<li>Personalization: the ability to adapt its functionality to increase its personal relevance to an individual or a category of individuals.</li>
</ol>
Personification
Personification refers to assigning personal traits to non-human agents, including physical appearance, and emotional states. Social characteristics related to personification include:
<ol>
<li>Identity: although chatbots lack the agency to choose what social group to which they want to belong, designers attribute identity to them, intentionally or not, when they define the way a chatbot talks or behaves;</li>
<li>Personality: personal traits that help to predict someone&#8217;s thoughts, feelings, and behaviors.</li>
</ol>
How to select social characterisctics for chatbots?
Chatbot tends to require more customized social traits in specific domains. In the education and customer service domains, for example, there are greater needs for manners and emotional intelligence, but chatbots may perform very differently guided with different goals. In addition, researchers have found that social characteristics can influence and promote each other.
These findings help provide insights into the compatability of specific characteristics with specific interaction scenarios which will help designers of AI-based chatbots make informed decisions when selecting a subset of characteristics.
<a href="https://www.tandfonline.com/doi/abs/10.1080/10447318.2020.1841438" target="_blank" rel="noopener">Read the paper on International Journal of Huma</a><a href="https://www.tandfonline.com/doi/abs/10.1080/10447318.2020.1841438" target="_blank" rel="noopener">n-Computer Interaction</a>

The eleven social characteristics necessary for Chatbots

[:en]The eleven social characteristics necessary for Chatbots[:zh]聊天机器人也要懂“人情世故”？一文了解Chatbot所需的11种社交特性[:]

In this bustling world, each of us journeys through life, carrying our own stories and emotions, in search of a sense of belonging amidst the flow of time. Music, the universal language of humanity, crosses cultural and geographical boundaries with its unique charm, soothing our souls. Among its many facets, melody has the most profound ability to touch our hearts. But how do these melodies navigate the depths of our soul, and how does our brain encode and understand them?
A recent study published in Science Advances offers insights into this question, presenting a complex and detailed map of neural encoding in the cerebral cortex as people listen to melodies. (This research, initially available as a preprint on bioRxiv, was highlighted by Edward Chang during the &#8220;Music and Healing: Scientific Exploration and Clinical Practice&#8221; symposium, with subsequent coverage by NextQuestion.)
<img loading="lazy" decoding="async" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1-1024x261.png" alt="" width="1024" height="261" class="aligncenter size-large wp-image-1856" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1-1024x261.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1-300x76.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1-768x195.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1.png 1297w" sizes="(max-width: 1024px) 100vw, 1024px" />
▷ Paper：Sankaran, Narayan, et al. &#8220;Encoding of melody in the human auditory cortex.&#8221; Science Advances 10.7 (2024): eadk0010. DOI: 10.1126/sciadv.adk0010
<h2>1. Fine-Grained Encoding of Musical Melody Dimensions in the Brain</h2>

To investigate how the brain encodes various dimensions of musical melodies, a research team led by Professor Edward Chang, a renowned neurosurgery expert from the University of California, San Francisco, developed a set of natural musical stimuli. This collection consisted of 208 short musical phrases arranged for different instruments, varying across three fundamental pitch-related dimensions: the absolute pitch of notes, the pitch change between adjacent notes, and musical expectancy based on preceding notes.
Employing high-density electrocorticography (ECoG), the researchers recorded neural physiological activity directly from the auditory cortex of eight participants. Analysis of high-frequency activity (HFA) revealed significant activity in the bilateral superior temporal gyrus (STG).
The team observed different neural response patterns when comparing notes of varying pitch (high vs. low), pitch changes (rising vs. falling), and expectancy (expected vs. unexpected). For instance, one electrode responded distinctively to different pitch ranges but not to pitch change or expectancy. Another electrode was sensitive to the direction of pitch changes, and a third electrode responded based on note expectancy. These findings suggest that various musical features are encoded independently by distinct neural populations within different areas of the auditory cortex.
<img loading="lazy" decoding="async" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1-1024x833.png" alt="" width="1024" height="833" class="aligncenter size-large wp-image-1855" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1-1024x833.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1-300x244.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1-768x625.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1.png 1268w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ Figure 1. Melody pitch, pitch changes, and expectancy influence STG activity during music listening. (A) Visual representation of three melodic features of a sample melody; (B) Electrodes from all participants overlaid on a standardized brain model. Colors represent the average induced high-frequency activity peak across all musical phrases. (C) Responses at three sample electrodes show distinct preferences for pitch (left column), pitch change (middle column), and expectancy (right column). Figure source: The paper.
To quantify how each electrode encoded these melody dimensions, the researchers employed temporal receptive field (TRF) modeling, which predicts neural activity based on a set of stimulus features. The results indicated that melodic and acoustic features significantly contributed to the variance in neural activity across electrodes, with the model accounting for an average of 70% of the explainable variance. The study also identified that specific melodic features are encoded independently by distinct subpopulations in the STG. These subpopulations extensively overlap along the STG&#8217;s posterior-anterior axis without forming anatomically discrete subareas.
Analysis of specific electrodes&#8217; sensitivity to melodic features provided evidence that different neural populations encode information about pitch, pitch changes, and expectancy in unique ways. For example, some electrodes were finely tuned to low pitches, others were particularly responsive to ascending pitch changes, and yet others were sensitive to notes of higher unexpectedness.
<h2>2. The Impact of Melodic Expectancy on Music-Selective Activity</h2>

To explore the impact of melodic expectancy on music-selective activity, researchers exposed eight participants to natural English sentences and the previously mentioned musical stimuli, recording their electroencephalographic (EEG) activity. By comparing the relative strength of responses to music and language, they developed a selectivity index (SI) ranging from -1 (indicating a preference for language) to +1 (indicating a preference for music). They discovered that a significant number of electrodes exhibited consistent and strong selective responses to musical phrases over spoken expressions.
These music-selective electrodes were found to be widely distributed and interspersed among other sound-responsive groups within the superior temporal gyrus (STG), indicating that music selectivity is not confined to any specialized auditory subregion of the brain. Further investigation showed that electrodes with a stronger response to music than to language almost exclusively processed melodic expectancy. Specifically, an electrode&#8217;s degree of melodic expectancy encoding was predictive of its music selectivity, while encoding for pitch or pitch changes did not systematically correlate with selectivity. This suggests that while the brain employs common neurons to assess pitch qualities in both speech and music, each modality has specific neurons dedicated to prediction.
Additionally, research into the encoding of natural language&#8217;s statistical structure revealed that phoneme-based expectancy, akin to melodic expectancy, is encoded by electrodes with language-selective responses in the STG. This implies that the brain encodes statistical information related to music and language through distinct neural substrates.
To determine whether music selectivity is driven exclusively by melodic expectancy rather than by low-level acoustic differences, researchers introduced a novel stimulus termed &#8216;melodic language.&#8217; This stimulus transformed the continuous intonation of language into discrete musical tones that aligned with melodies while preserving other acoustic features of spoken language. Results indicated that melodic language elicited strong responses from music-selective electrodes, even when acoustic features closely resembled conventional language, and the strength of this response positively correlated with participants&#8217; perception of melody.
These findings collectively underscore the conclusion that music-selective activity, particularly when processing music and melodic language, predominantly reflects the encoding of melodic expectancy, rather than being based on low-level acoustic characteristics.
<h2>3. Encoding of Melody Pitch and Pitch Changes Shared Between Music and Speech</h2>

Researchers identified acoustic dimensions akin to melody pitch and pitch changes present in both music and language by analyzing the pitch contour of sentences and calculating the pitch changes between adjacent syllables. Despite the differences in the distribution of pitch and pitch changes across these domains, a substantial overlap was found, showcasing a shared foundation in these fundamental acoustic properties between music and language.
The study compared the unique variance explained by each feature in the temporal receptive field (TRF) model of language-induced activity, revealing a strong correlation in the encoding intensity by electrodes for both pitch and pitch changes across the two stimulus domains. This suggests that neural populations attuned to specific aspects of melody are likely adept at encoding the corresponding dimension in language as well.
To delve into how these neural populations employ a cross-domain, universal neural code for information representation, researchers analyzed tuning curves for the common range of pitch and pitch changes in music and language. The high correlation between tuning curves for both domains underscores a deeply conserved neural encoding mechanism.
Moreover, the study utilized a linear classifier to decode the direction of pitch or pitch changes based on aggregated electrode activity from participants. This classifier, trained and tested within both musical and linguistic contexts, underscored that, unlike the encoding of expectancy in musical melodies, basic musical attributes like pitch and pitch changes are encapsulated within a universal neural code also relevant to language.
<h2>4. Conclusion</h2>

At the nexus of music and language stands a bridge constructed by neuroscience. This bridge not only links two seemingly disparate domains but also reveals a profound unity in how the human brain processes sound information.
As we listen to music, our brain engages with the pitch of notes through two sets of neurons that also tune into the tonal rhythms of speech. It anticipates the next note with a specialized set of neurons. The superior temporal gyrus (STG) serves not just as a sanctuary for music; it is a complex information processing hub that distinguishes subtle pitch variations, traces the movement of notes as they jump or slide, and even anticipates the emergence of subsequent notes.
The elements of pitch and pitch changes, foundational to musical melody, are not regarded by our brain as exclusive to either music or language. They form a shared neural code, whether in the enjoyment of a melodic piece or in the listening to daily conversation. These essential sound properties are encoded and understood by our brain in a similar manner. Indeed, like the continuous flow of a river merging with the sea, music and language converge in the vast expanse of human cognitive processing.

[:en]In this bustling world, each of us journeys through life, carrying our own stories and emotions, in search of a sense of belonging amidst the flow of time. Music, the universal language of humanity, crosses cultural and geographical boundaries with its unique charm, soothing our souls. Among its many facets, melody has the most profound ability to touch our hearts. But how do these melodies navigate the depths of our soul, and how does our brain encode and understand them?

A recent study published in Science Advances offers insights into this question, presenting a complex and detailed map of neural encoding in the cerebral cortex as people listen to melodies. (This research, initially available as a preprint on bioRxiv, was highlighted by Edward Chang during the "Music and Healing: Scientific Exploration and Clinical Practice" symposium, with subsequent coverage by NextQuestion.)

<img src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1-1024x261.png" alt="" width="1024" height="261" class="aligncenter size-large wp-image-1856" />

▷ Paper：Sankaran, Narayan, et al. "Encoding of melody in the human auditory cortex." Science Advances 10.7 (2024): eadk0010. DOI: 10.1126/sciadv.adk0010

<h2>1. Fine-Grained Encoding of Musical Melody Dimensions in the Brain</h2>

To investigate how the brain encodes various dimensions of musical melodies, a research team led by Professor Edward Chang, a renowned neurosurgery expert from the University of California, San Francisco, developed a set of natural musical stimuli. This collection consisted of 208 short musical phrases arranged for different instruments, varying across three fundamental pitch-related dimensions: the absolute pitch of notes, the pitch change between adjacent notes, and musical expectancy based on preceding notes.

Employing high-density electrocorticography (ECoG), the researchers recorded neural physiological activity directly from the auditory cortex of eight participants. Analysis of high-frequency activity (HFA) revealed significant activity in the bilateral superior temporal gyrus (STG).

The team observed different neural response patterns when comparing notes of varying pitch (high vs. low), pitch changes (rising vs. falling), and expectancy (expected vs. unexpected). For instance, one electrode responded distinctively to different pitch ranges but not to pitch change or expectancy. Another electrode was sensitive to the direction of pitch changes, and a third electrode responded based on note expectancy. These findings suggest that various musical features are encoded independently by distinct neural populations within different areas of the auditory cortex.

<img src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture5-1-1024x833.png" alt="" width="1024" height="833" class="aligncenter size-large wp-image-1855" />
▷ Figure 1. Melody pitch, pitch changes, and expectancy influence STG activity during music listening. (A) Visual representation of three melodic features of a sample melody; (B) Electrodes from all participants overlaid on a standardized brain model. Colors represent the average induced high-frequency activity peak across all musical phrases. (C) Responses at three sample electrodes show distinct preferences for pitch (left column), pitch change (middle column), and expectancy (right column). Figure source: The paper.

To quantify how each electrode encoded these melody dimensions, the researchers employed temporal receptive field (TRF) modeling, which predicts neural activity based on a set of stimulus features. The results indicated that melodic and acoustic features significantly contributed to the variance in neural activity across electrodes, with the model accounting for an average of 70% of the explainable variance. The study also identified that specific melodic features are encoded independently by distinct subpopulations in the STG. These subpopulations extensively overlap along the STG's posterior-anterior axis without forming anatomically discrete subareas.

Analysis of specific electrodes' sensitivity to melodic features provided evidence that different neural populations encode information about pitch, pitch changes, and expectancy in unique ways. For example, some electrodes were finely tuned to low pitches, others were particularly responsive to ascending pitch changes, and yet others were sensitive to notes of higher unexpectedness.

<h2>2. The Impact of Melodic Expectancy on Music-Selective Activity</h2>

To explore the impact of melodic expectancy on music-selective activity, researchers exposed eight participants to natural English sentences and the previously mentioned musical stimuli, recording their electroencephalographic (EEG) activity. By comparing the relative strength of responses to music and language, they developed a selectivity index (SI) ranging from -1 (indicating a preference for language) to +1 (indicating a preference for music). They discovered that a significant number of electrodes exhibited consistent and strong selective responses to musical phrases over spoken expressions.

These music-selective electrodes were found to be widely distributed and interspersed among other sound-responsive groups within the superior temporal gyrus (STG), indicating that music selectivity is not confined to any specialized auditory subregion of the brain. Further investigation showed that electrodes with a stronger response to music than to language almost exclusively processed melodic expectancy. Specifically, an electrode's degree of melodic expectancy encoding was predictive of its music selectivity, while encoding for pitch or pitch changes did not systematically correlate with selectivity. This suggests that while the brain employs common neurons to assess pitch qualities in both speech and music, each modality has specific neurons dedicated to prediction.

Additionally, research into the encoding of natural language's statistical structure revealed that phoneme-based expectancy, akin to melodic expectancy, is encoded by electrodes with language-selective responses in the STG. This implies that the brain encodes statistical information related to music and language through distinct neural substrates.

To determine whether music selectivity is driven exclusively by melodic expectancy rather than by low-level acoustic differences, researchers introduced a novel stimulus termed 'melodic language.' This stimulus transformed the continuous intonation of language into discrete musical tones that aligned with melodies while preserving other acoustic features of spoken language. Results indicated that melodic language elicited strong responses from music-selective electrodes, even when acoustic features closely resembled conventional language, and the strength of this response positively correlated with participants' perception of melody.

These findings collectively underscore the conclusion that music-selective activity, particularly when processing music and melodic language, predominantly reflects the encoding of melodic expectancy, rather than being based on low-level acoustic characteristics.

<h2>3. Encoding of Melody Pitch and Pitch Changes Shared Between Music and Speech</h2>

Researchers identified acoustic dimensions akin to melody pitch and pitch changes present in both music and language by analyzing the pitch contour of sentences and calculating the pitch changes between adjacent syllables. Despite the differences in the distribution of pitch and pitch changes across these domains, a substantial overlap was found, showcasing a shared foundation in these fundamental acoustic properties between music and language.

The study compared the unique variance explained by each feature in the temporal receptive field (TRF) model of language-induced activity, revealing a strong correlation in the encoding intensity by electrodes for both pitch and pitch changes across the two stimulus domains. This suggests that neural populations attuned to specific aspects of melody are likely adept at encoding the corresponding dimension in language as well.

To delve into how these neural populations employ a cross-domain, universal neural code for information representation, researchers analyzed tuning curves for the common range of pitch and pitch changes in music and language. The high correlation between tuning curves for both domains underscores a deeply conserved neural encoding mechanism.

Moreover, the study utilized a linear classifier to decode the direction of pitch or pitch changes based on aggregated electrode activity from participants. This classifier, trained and tested within both musical and linguistic contexts, underscored that, unlike the encoding of expectancy in musical melodies, basic musical attributes like pitch and pitch changes are encapsulated within a universal neural code also relevant to language.

<h2>4. Conclusion</h2>

At the nexus of music and language stands a bridge constructed by neuroscience. This bridge not only links two seemingly disparate domains but also reveals a profound unity in how the human brain processes sound information.

As we listen to music, our brain engages with the pitch of notes through two sets of neurons that also tune into the tonal rhythms of speech. It anticipates the next note with a specialized set of neurons. The superior temporal gyrus (STG) serves not just as a sanctuary for music; it is a complex information processing hub that distinguishes subtle pitch variations, traces the movement of notes as they jump or slide, and even anticipates the emergence of subsequent notes.

The elements of pitch and pitch changes, foundational to musical melody, are not regarded by our brain as exclusive to either music or language. They form a shared neural code, whether in the enjoyment of a melodic piece or in the listening to daily conversation. These essential sound properties are encoded and understood by our brain in a similar manner. Indeed, like the continuous flow of a river merging with the sea, music and language converge in the vast expanse of human cognitive processing.
[:]

What Happens in Our Brain When We Listen to Music?

[:en]What Happens in Our Brain When We Listen to Music?[:]

Is thinking the spark of inspiration or the burden of the mind? Recall the night before the college entrance exam, where we all struggled with math problems: faced with the last major question on the math exam, scratching our heads, drafting and redrafting, pen lifted and then set down, repeatedly. In that moment, thinking undoubtedly became a heavy burden, like a mountain too daunting to climb.
Yet, in everyday life, many problems seem to solve themselves effortlessly, akin to simple arithmetic. Even for geometric fill-in-the-blank questions, practiced countless times and etched into our hearts, we often glimpse the answer instantly. In such instances, thinking appears so effortless, as if without the need to think, the answer naturally presents itself before our eyes.
For decades, psychologists have embraced the &#8220;dual-process theory&#8221; to distinguish between these two types of thinking. One type, swift and effortless as lightning, driven by intuition, is called System 1. The other, akin to climbing steep mountains, demanding deliberation, slow and challenging, is referred to as System 2.
You might not know this theory by name, but you are likely familiar with the book &#8220;Thinking, Fast and Slow&#8221; by Daniel Kahneman, the 2002 Nobel Prize in Economics laureate. The dual-process theory has been widely applied across various research areas, including judgment and decision-making, morality and human cooperation, and social cognition. Beyond academia, the theory also enjoys popularity among the public, influencing numerous policy formulations.
However, even such an influential theory faces its criticisms and challenges. Recently, researcher Wim De Neys from the University of Paris published an extensive article in Behavioral and Brain Sciences, discussing the limitations of the dual-process theory. His proposed new model, building on the dual-process framework, has sparked lively discussions among scholars.
<img loading="lazy" decoding="async" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1.png" alt="" width="1013" height="318" class="aligncenter size-full wp-image-1851" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1.png 1013w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1-300x94.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1-768x241.png 768w" sizes="(max-width: 1013px) 100vw, 1013px" /> 
▷ De Neys, Wim. &#8220;Advancing theorizing about fast-and-slow thinking.&#8221; Behavioral and Brain Sciences 46 (2023): e111.
<h2>1. Born from Conflict: The Dual-process</h2>

The dual-process theory, initially proposed by Peter Wason and Jonathan Evans and later refined by Daniel Kahneman, delineates human thought into two distinct modes: System 1 and System 2. System 1, the intuitive system, operates quickly and automatically, driven by emotions and experiences, and is challenging to control or change. In contrast, System 2, the deliberative system, is slow, rational, and conscious, requiring more resources but less susceptible to mistakes. The theory posits that these systems are fundamentally different and exclusive, leading to divergent responses, with System 2 capable of generating solutions beyond System 1&#8217;s reach. Empirical research across various disciplines supports this theory.
The existence of System 1 explains why people&#8217;s choices often defy logic and probability. Consider choosing between two jars for a chance to win a prize by drawing a red marble: Jar A, with 10 marbles including one red, and Jar B, with 100 marbles including nine red. Despite the better odds in Jar A (1/10) compared to Jar B (9/100) as per probability theory (System 2), the greater number of red marbles in Jar B (9 > 1) often makes it the more appealing choice due to intuitive heuristics (System 1).
The dual-process theory frequently examines conflict scenarios, such as the marble problem, where specific cues trigger intuitive responses. Due to a preference for minimizing cognitive effort, individuals may rely on intuition rather than engage in thorough, slow thinking. This tendency can change with sufficient resources and motivation, leading to deeper consideration. For instance, students may not scrutinize all options when an answer seems obvious during practice but will do so during critical exams to ensure accuracy.
Researchers in various fields use conflict scenarios to contrast the two systems, as seen in moral reasoning studies with trolley problems. Here, deontological reactions (i.e., avoiding harm) represent fast, intuitive System 1 responses, while utilitarian choices (i.e., maximizing overall well-being) reflect the more analytical System 2, involving complex decision-making processes.
Such conflict scenarios often categorize responses aligned with System 1 as &#8220;biases&#8221; and those with System 2 as &#8220;rational,&#8221; underscoring the theory&#8217;s assumption of exclusivity. This perspective is prevalent in research on heuristics and cognitive biases. Nonetheless, it&#8217;s crucial to recognize that not all System 1 responses are inherently biased; for straightforward queries like &#8220;What is 1 + 1?&#8221;, accurate answers can emerge swiftly from System 1.
<h2>2. The Exclusivity Hypothesis: The Dual-process</h2>

As previously discussed, System 2 is often activated when resources are abundant or motivation is high. Given its reputation for being more &#8220;rational&#8221; and producing responses beyond the reach of System 1, the question arises: when do people actually rely on System 2 for thinking? The cognitive load paradigm has been extensively used in research to investigate the transition between these two systems.
If engaging System 2 for thought demands more time and cognitive resources, then under the pressure of time constraints or cognitive overload, individuals will default to relying on the quick, instinctive reactions of System 1 due to a lack of resources. Numerous studies have supported this hypothesis; under restrictive conditions, people are less inclined to engage in deep thinking with System 2, lending credence to the dual-process&#8217; exclusivity hypothesis.
However, recent replication efforts and meta-analyses have cast doubt on the substantial impact of cognitive constraints. For instance, Lawson et al. (2020) observed that, in logic reasoning tasks, the difference in choice due to cognitive constraints was a mere 9.4%, and even under such conditions, more than 50% of decisions aligned with System 2 responses. This finding suggests that people are capable of making rational choices in line with System 2 even when conditions are not conducive to in-depth thinking. This challenges the notion that System 1&#8217;s intuitive responses cannot mirror those of deliberative thinking, thus questioning the exclusivity hypothesis.
The dual-response paradigm provides further evidence against this exclusivity. Thompson et al. (2011) had participants answer a question twice: initially, they were to respond quickly with the first answer that came to mind, and subsequently, they were allowed time to deliberate before giving a final answer. To ensure the initial response was intuitive (System 1), the first reply was given under time pressure or cognitive load. According to the exclusivity hypothesis, a deliberative, System 2-driven final response should correct an initial System 1-based intuition. However, empirical results have not supported this hypothesis. For example, in the classic logic problem involving the cost of a bat and a ball, many individuals who arrived at the correct answer ($0.05) did so initially, on their first, intuitive attempt.
The conflict detection paradigm further challenges the exclusivity hypothesis. In studies using this approach, participants are divided into groups faced with tasks that either present a potential conflict between the two systems or do not. Findings reveal that even in tasks where System 1 drives the initial response, participants can detect conflicts, taking longer and expressing less confidence in their answers (e.g., Šrol &#038; De Neys, 2021). This demonstrates that, even when relying on quick intuitive decisions, individuals can still discern the complexity and potential inaccuracies in their choices.
Do the dual-process truly operate exclusively of each other? Is System 1 incapable of producing the kinds of responses attributed to System 2? Given the evidence from various research paradigms, it&#8217;s challenging to assert the inferiority of System 1. It appears more plausible that System 1, typically associated with intuition, is also capable of yielding deliberative-type responses.
<h2>3. The Timing for System 2 Engagement</h2>

Given System 1&#8217;s susceptibility to biased responses, the question arises: How do individuals transition to using System 2? This shift has been the subject of various hypotheses.
&#8211; System 2&#8217;s Conflict Monitoring Ability
The dual-processes are often conceptualized as a serial processing architecture with default-interventionist characteristics, where we predominantly rely on System 1&#8217;s unconscious processing. System 2 is invoked for in-depth analysis when conflicts or challenges are detected. The classical view suggests that System 2 can identify conflicts arising from System 1, intervening as necessary. However, this raises a question: If System 2 preemptively recognizes conflicts, implying its activation prior to conflict, how is it &#8220;reactivated&#8221; for resolution?
&#8211; Low-Energy Deliberation
To refine this, some researchers propose that System 2 consists of multiple components, where conflict monitoring requires only a fraction of System 2&#8217;s capacity, utilizing less energy. Yet, this leads to another conundrum: If the deliberative functions of System 2 are dormant, and a low-energy segment is merely monitoring System 1&#8217;s outputs, what exactly triggers conflict detection within this low-energy segment?
&#8211; System 3
Considering the possibility that System 2 may not concurrently monitor and respond, could these monitoring duties be attributed to a hypothetical System 3? This notion, while offering a solution for conflict monitoring, still suggests that System 2 has formulated a response, undermining the concept of exclusive System 1 activity.
&#8211; Parallel Processing
Another view posits that Systems 1 and 2 can be simultaneously active, obviating a strict switch mechanism. However, this model, featuring a perpetually active, resource-intensive System 2 alongside a System 1 awaiting conflict verification, seems economically inefficient from a cognitive resource standpoint.
The traditional explanations for system transition exhibit limitations. Following an exhaustive review of dual-process theory research, De Neys has proposed a new model that transcends the exclusivity assumption, providing a more nuanced understanding of the mechanisms behind system switching.
<h2>4. Intuitive Responses &#038; Uncertainty Monitoring</h2>

As previously discussed, System 1 is also capable of generating &#8220;correct&#8221; responses. On this foundation, De Neys has restructured the dual-process theory, as depicted below.
<img loading="lazy" decoding="async" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1.png" alt="" width="878" height="968" class="aligncenter size-full wp-image-1850" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1.png 878w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1-272x300.png 272w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1-768x847.png 768w" sizes="(max-width: 878px) 100vw, 878px" />
▷ Figure: This model encompasses four core components: System 1 (intuitive activation), uncertainty monitoring, System 2 (deliberation), and the feedback loop. In 1, I1 and I2 represent distinct intuitions, with the scatter plot illustrating the changes in activation strength of different intuitions over time. In 2, the blue curve depicts how uncertainty fluctuates with events, and d represents the uncertainty threshold. Individuals engage System 2 for deeper thinking only when uncertainty surpasses threshold d. The interventions of System 2 (such as modifying existing intuitions or crafting new responses) ultimately feed back into System 1 through the feedback loop. Figure source: De Neys (2023)
The initial phase involves System 1 generating a variety of intuitions, with different stimuli triggering specific intuitive responses. The activation level of these intuitions changes over time, peaking in strength based on the association with the cue that elicited them. In certain reasoning tasks, cues could include aspects like the number or proportion of red marbles.
The subsequent phase is uncertainty monitoring, where System 1 constantly assesses the degree of uncertainty by evaluating the variance in activation levels among various intuitions. De Neys offers a straightforward calculation for this: U = 1 – |I1 – I2|, where U denotes uncertainty, and I1 and I2 represent the activation strengths of two different intuitions, each ranging from 0 to 1. The lesser the difference in activation strengths between the two intuitions, the greater the uncertainty. De Neys posits that when uncertainty exceeds a specific threshold (d), it triggers the utilization of System 2 for in-depth analysis.
In its role, System 2 can undertake various functions within the thinking process, such as inhibiting responses, generating new ones, or rationalization. When faced with heightened uncertainty, System 2 engages, deliberating over the activated intuitions, either by tweaking the activation strengths of already generated intuitive responses or by formulating new ones distinct from existing intuitions. This process elucidates why employing System 2 does not invariably lead to &#8220;correct&#8221; or &#8220;rational&#8221; outcomes, corroborating findings from earlier research.
De Neys&#8217;s model culminates with the feedback loop. The adjustments made by System 2 to the activation strengths of different intuitions in System 1, or the new responses it generates, modify the level of uncertainty. Once the uncertainty dips below the threshold d, System 2 disengages, allowing the thought process to conclude predominantly with System 1.
In essence, De Neys clarifies the distinct roles of System 1 and System 2 in thinking. System 1 is tasked with generating and representing various intuitive responses and calculating uncertainty based on these intuitions&#8217; activation levels. The degree of uncertainty dictates the engagement of System 2; elevated uncertainty prompts System 2 to further process the intuitive responses from System 1 (e.g., adjustments, new response generation).
To contextualize De Neys&#8217;s model, consider a real-life scenario: lunchtime at a cafeteria with a choice between ice cream and cupcakes for dessert. Xiao Ming, who prefers ice cream, quickly decides based solely on System 1. Conversely, Xiao Gang, who likes both desserts, faces uncertainty, necessitating more deliberation (System 2 engagement). During this process, he might recall past experiences or consider recommendations, adjusting his preferences (feedback loop). Ultimately, remembering a previous stomachache from ice cream decisively lowers his uncertainty, leading him to choose the cupcake.
<h2>5. Conclusion</h2>

The critique of the dual-process theory by De Neys, along with his proposed new model, has attracted significant attention and commentary from numerous researchers. Currently, this model is still in the early stages of conceptual development, with many areas awaiting refinement.
For example, Ackerman and Morsanyi have raised concerns that De Neys&#8217;s model may not fully incorporate findings from metacognition, particularly the role of confidence in determining when to cease thinking. Baron has suggested that individual differences could influence the decision to stop deliberation. He posits that individuals who trust their intuition might halt after System 1&#8217;s response; those who are open-minded might pause only after exploring all possibilities; and individuals who are averse to uncertainty could seek confirmatory evidence for their intuition more diligently, aiming to minimize the uncertainty associated with further contemplation.
These discussions lead to a more fundamental inquiry: Is the dual-process theory necessary? Or, more pointedly, do humans indeed operate with dual-process? Many of De Neys&#8217;s criticisms also challenge this core premise.
Frankish proposes an extension of De Neys&#8217;s ideas, suggesting that System 2 might actually be a product of System 1. In his view, System 2 is not independent but rather a consequence of System 1&#8217;s conscious, sustained engagement with specific stimuli, acting as a &#8220;virtual&#8221; system. When confronted with complex tasks, we essentially decompose them into simpler subtasks for System 1 to process. For instance, in multiplying two-digit numbers like &#8220;11 × 84 = ?&#8221;, we break it down into manageable parts: &#8220;11 × 4 = 44&#8221;, &#8220;11 × 80 = 880&#8221;, and &#8220;44 + 880 = 924&#8221;. For an adult, each component is simpler than the direct calculation, facilitating rapid solutions.
Indeed, many problems that initially seem to require intensive thought can become intuitive through continual learning, showcasing the dynamic interaction between System 2 and System 1. For instance, once children learn multiplication tables, they no longer need to resort to addition for such problems, instead instantly recalling facts like &#8220;eight eights are sixty-four&#8221;, &#8220;nine nines are eighty-one&#8221;.
This reflection prompts us to consider the original purpose behind the proposal of the dual-process theory. As Tinghög and others have remarked, was it introduced to generate testable predictions? Perhaps not; it appears researchers regard the dual-process theory more as a theoretical framework rather than a predictive model. Despite the level of empirical support, the theory offers a valuable starting point, enabling researchers to formulate more precise questions and develop more sophisticated theories.
Nevertheless, the dual-process theory may still represent a useful heuristic, given humanity&#8217;s tendency to categorize the world in binaries: good versus bad, black versus white, fast versus slow, self versus others. With this in mind, does the dual-process theory enhance our understanding of human thought? Perhaps answering this question itself necessitates the engagement of System 2.

[:en]Is thinking the spark of inspiration or the burden of the mind? Recall the night before the college entrance exam, where we all struggled with math problems: faced with the last major question on the math exam, scratching our heads, drafting and redrafting, pen lifted and then set down, repeatedly. In that moment, thinking undoubtedly became a heavy burden, like a mountain too daunting to climb.

Yet, in everyday life, many problems seem to solve themselves effortlessly, akin to simple arithmetic. Even for geometric fill-in-the-blank questions, practiced countless times and etched into our hearts, we often glimpse the answer instantly. In such instances, thinking appears so effortless, as if without the need to think, the answer naturally presents itself before our eyes.

For decades, psychologists have embraced the "dual-process theory" to distinguish between these two types of thinking. One type, swift and effortless as lightning, driven by intuition, is called System 1. The other, akin to climbing steep mountains, demanding deliberation, slow and challenging, is referred to as System 2.

You might not know this theory by name, but you are likely familiar with the book "Thinking, Fast and Slow" by Daniel Kahneman, the 2002 Nobel Prize in Economics laureate. The dual-process theory has been widely applied across various research areas, including judgment and decision-making, morality and human cooperation, and social cognition. Beyond academia, the theory also enjoys popularity among the public, influencing numerous policy formulations.

However, even such an influential theory faces its criticisms and challenges. Recently, researcher Wim De Neys from the University of Paris published an extensive article in Behavioral and Brain Sciences, discussing the limitations of the dual-process theory. His proposed new model, building on the dual-process framework, has sparked lively discussions among scholars.

<img src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1.png" alt="" width="1013" height="318" class="aligncenter size-full wp-image-1851" />
▷ De Neys, Wim. "Advancing theorizing about fast-and-slow thinking." Behavioral and Brain Sciences 46 (2023): e111.

<h2>1. Born from Conflict: The Dual-process</h2>

The dual-process theory, initially proposed by Peter Wason and Jonathan Evans and later refined by Daniel Kahneman, delineates human thought into two distinct modes: System 1 and System 2. System 1, the intuitive system, operates quickly and automatically, driven by emotions and experiences, and is challenging to control or change. In contrast, System 2, the deliberative system, is slow, rational, and conscious, requiring more resources but less susceptible to mistakes. The theory posits that these systems are fundamentally different and exclusive, leading to divergent responses, with System 2 capable of generating solutions beyond System 1's reach. Empirical research across various disciplines supports this theory.

The existence of System 1 explains why people's choices often defy logic and probability. Consider choosing between two jars for a chance to win a prize by drawing a red marble: Jar A, with 10 marbles including one red, and Jar B, with 100 marbles including nine red. Despite the better odds in Jar A (1/10) compared to Jar B (9/100) as per probability theory (System 2), the greater number of red marbles in Jar B (9 > 1) often makes it the more appealing choice due to intuitive heuristics (System 1).

The dual-process theory frequently examines conflict scenarios, such as the marble problem, where specific cues trigger intuitive responses. Due to a preference for minimizing cognitive effort, individuals may rely on intuition rather than engage in thorough, slow thinking. This tendency can change with sufficient resources and motivation, leading to deeper consideration. For instance, students may not scrutinize all options when an answer seems obvious during practice but will do so during critical exams to ensure accuracy.

Researchers in various fields use conflict scenarios to contrast the two systems, as seen in moral reasoning studies with trolley problems. Here, deontological reactions (i.e., avoiding harm) represent fast, intuitive System 1 responses, while utilitarian choices (i.e., maximizing overall well-being) reflect the more analytical System 2, involving complex decision-making processes.

Such conflict scenarios often categorize responses aligned with System 1 as "biases" and those with System 2 as "rational," underscoring the theory's assumption of exclusivity. This perspective is prevalent in research on heuristics and cognitive biases. Nonetheless, it's crucial to recognize that not all System 1 responses are inherently biased; for straightforward queries like "What is 1 + 1?", accurate answers can emerge swiftly from System 1.

<h2>2. The Exclusivity Hypothesis: The Dual-process</h2>

As previously discussed, System 2 is often activated when resources are abundant or motivation is high. Given its reputation for being more "rational" and producing responses beyond the reach of System 1, the question arises: when do people actually rely on System 2 for thinking? The cognitive load paradigm has been extensively used in research to investigate the transition between these two systems.

If engaging System 2 for thought demands more time and cognitive resources, then under the pressure of time constraints or cognitive overload, individuals will default to relying on the quick, instinctive reactions of System 1 due to a lack of resources. Numerous studies have supported this hypothesis; under restrictive conditions, people are less inclined to engage in deep thinking with System 2, lending credence to the dual-process' exclusivity hypothesis.

However, recent replication efforts and meta-analyses have cast doubt on the substantial impact of cognitive constraints. For instance, Lawson et al. (2020) observed that, in logic reasoning tasks, the difference in choice due to cognitive constraints was a mere 9.4%, and even under such conditions, more than 50% of decisions aligned with System 2 responses. This finding suggests that people are capable of making rational choices in line with System 2 even when conditions are not conducive to in-depth thinking. This challenges the notion that System 1's intuitive responses cannot mirror those of deliberative thinking, thus questioning the exclusivity hypothesis.

The dual-response paradigm provides further evidence against this exclusivity. Thompson et al. (2011) had participants answer a question twice: initially, they were to respond quickly with the first answer that came to mind, and subsequently, they were allowed time to deliberate before giving a final answer. To ensure the initial response was intuitive (System 1), the first reply was given under time pressure or cognitive load. According to the exclusivity hypothesis, a deliberative, System 2-driven final response should correct an initial System 1-based intuition. However, empirical results have not supported this hypothesis. For example, in the classic logic problem involving the cost of a bat and a ball, many individuals who arrived at the correct answer ($0.05) did so initially, on their first, intuitive attempt.

The conflict detection paradigm further challenges the exclusivity hypothesis. In studies using this approach, participants are divided into groups faced with tasks that either present a potential conflict between the two systems or do not. Findings reveal that even in tasks where System 1 drives the initial response, participants can detect conflicts, taking longer and expressing less confidence in their answers (e.g., Šrol & De Neys, 2021). This demonstrates that, even when relying on quick intuitive decisions, individuals can still discern the complexity and potential inaccuracies in their choices.

Do the dual-process truly operate exclusively of each other? Is System 1 incapable of producing the kinds of responses attributed to System 2? Given the evidence from various research paradigms, it's challenging to assert the inferiority of System 1. It appears more plausible that System 1, typically associated with intuition, is also capable of yielding deliberative-type responses.

<h2>3. The Timing for System 2 Engagement</h2>

Given System 1's susceptibility to biased responses, the question arises: How do individuals transition to using System 2? This shift has been the subject of various hypotheses.

- System 2's Conflict Monitoring Ability

The dual-processes are often conceptualized as a serial processing architecture with default-interventionist characteristics, where we predominantly rely on System 1's unconscious processing. System 2 is invoked for in-depth analysis when conflicts or challenges are detected. The classical view suggests that System 2 can identify conflicts arising from System 1, intervening as necessary. However, this raises a question: If System 2 preemptively recognizes conflicts, implying its activation prior to conflict, how is it "reactivated" for resolution?

- Low-Energy Deliberation

To refine this, some researchers propose that System 2 consists of multiple components, where conflict monitoring requires only a fraction of System 2's capacity, utilizing less energy. Yet, this leads to another conundrum: If the deliberative functions of System 2 are dormant, and a low-energy segment is merely monitoring System 1's outputs, what exactly triggers conflict detection within this low-energy segment?

- System 3

Considering the possibility that System 2 may not concurrently monitor and respond, could these monitoring duties be attributed to a hypothetical System 3? This notion, while offering a solution for conflict monitoring, still suggests that System 2 has formulated a response, undermining the concept of exclusive System 1 activity.

- Parallel Processing

Another view posits that Systems 1 and 2 can be simultaneously active, obviating a strict switch mechanism. However, this model, featuring a perpetually active, resource-intensive System 2 alongside a System 1 awaiting conflict verification, seems economically inefficient from a cognitive resource standpoint.

The traditional explanations for system transition exhibit limitations. Following an exhaustive review of dual-process theory research, De Neys has proposed a new model that transcends the exclusivity assumption, providing a more nuanced understanding of the mechanisms behind system switching.

<h2>4. Intuitive Responses & Uncertainty Monitoring</h2>

As previously discussed, System 1 is also capable of generating "correct" responses. On this foundation, De Neys has restructured the dual-process theory, as depicted below.

<img src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1.png" alt="" width="878" height="968" class="aligncenter size-full wp-image-1850" />

▷ Figure: This model encompasses four core components: System 1 (intuitive activation), uncertainty monitoring, System 2 (deliberation), and the feedback loop. In 1, I1 and I2 represent distinct intuitions, with the scatter plot illustrating the changes in activation strength of different intuitions over time. In 2, the blue curve depicts how uncertainty fluctuates with events, and d represents the uncertainty threshold. Individuals engage System 2 for deeper thinking only when uncertainty surpasses threshold d. The interventions of System 2 (such as modifying existing intuitions or crafting new responses) ultimately feed back into System 1 through the feedback loop. Figure source: De Neys (2023)

The initial phase involves System 1 generating a variety of intuitions, with different stimuli triggering specific intuitive responses. The activation level of these intuitions changes over time, peaking in strength based on the association with the cue that elicited them. In certain reasoning tasks, cues could include aspects like the number or proportion of red marbles.

The subsequent phase is uncertainty monitoring, where System 1 constantly assesses the degree of uncertainty by evaluating the variance in activation levels among various intuitions. De Neys offers a straightforward calculation for this: U = 1 – |I1 – I2|, where U denotes uncertainty, and I1 and I2 represent the activation strengths of two different intuitions, each ranging from 0 to 1. The lesser the difference in activation strengths between the two intuitions, the greater the uncertainty. De Neys posits that when uncertainty exceeds a specific threshold (d), it triggers the utilization of System 2 for in-depth analysis.

In its role, System 2 can undertake various functions within the thinking process, such as inhibiting responses, generating new ones, or rationalization. When faced with heightened uncertainty, System 2 engages, deliberating over the activated intuitions, either by tweaking the activation strengths of already generated intuitive responses or by formulating new ones distinct from existing intuitions. This process elucidates why employing System 2 does not invariably lead to "correct" or "rational" outcomes, corroborating findings from earlier research.

De Neys's model culminates with the feedback loop. The adjustments made by System 2 to the activation strengths of different intuitions in System 1, or the new responses it generates, modify the level of uncertainty. Once the uncertainty dips below the threshold d, System 2 disengages, allowing the thought process to conclude predominantly with System 1.

In essence, De Neys clarifies the distinct roles of System 1 and System 2 in thinking. System 1 is tasked with generating and representing various intuitive responses and calculating uncertainty based on these intuitions' activation levels. The degree of uncertainty dictates the engagement of System 2; elevated uncertainty prompts System 2 to further process the intuitive responses from System 1 (e.g., adjustments, new response generation).

To contextualize De Neys's model, consider a real-life scenario: lunchtime at a cafeteria with a choice between ice cream and cupcakes for dessert. Xiao Ming, who prefers ice cream, quickly decides based solely on System 1. Conversely, Xiao Gang, who likes both desserts, faces uncertainty, necessitating more deliberation (System 2 engagement). During this process, he might recall past experiences or consider recommendations, adjusting his preferences (feedback loop). Ultimately, remembering a previous stomachache from ice cream decisively lowers his uncertainty, leading him to choose the cupcake.

<h2>5. Conclusion</h2>

The critique of the dual-process theory by De Neys, along with his proposed new model, has attracted significant attention and commentary from numerous researchers. Currently, this model is still in the early stages of conceptual development, with many areas awaiting refinement.

For example, Ackerman and Morsanyi have raised concerns that De Neys's model may not fully incorporate findings from metacognition, particularly the role of confidence in determining when to cease thinking. Baron has suggested that individual differences could influence the decision to stop deliberation. He posits that individuals who trust their intuition might halt after System 1's response; those who are open-minded might pause only after exploring all possibilities; and individuals who are averse to uncertainty could seek confirmatory evidence for their intuition more diligently, aiming to minimize the uncertainty associated with further contemplation.

These discussions lead to a more fundamental inquiry: Is the dual-process theory necessary? Or, more pointedly, do humans indeed operate with dual-process? Many of De Neys's criticisms also challenge this core premise.

Frankish proposes an extension of De Neys's ideas, suggesting that System 2 might actually be a product of System 1. In his view, System 2 is not independent but rather a consequence of System 1's conscious, sustained engagement with specific stimuli, acting as a "virtual" system. When confronted with complex tasks, we essentially decompose them into simpler subtasks for System 1 to process. For instance, in multiplying two-digit numbers like "11 × 84 = ?", we break it down into manageable parts: "11 × 4 = 44", "11 × 80 = 880", and "44 + 880 = 924". For an adult, each component is simpler than the direct calculation, facilitating rapid solutions.

Indeed, many problems that initially seem to require intensive thought can become intuitive through continual learning, showcasing the dynamic interaction between System 2 and System 1. For instance, once children learn multiplication tables, they no longer need to resort to addition for such problems, instead instantly recalling facts like "eight eights are sixty-four", "nine nines are eighty-one".

This reflection prompts us to consider the original purpose behind the proposal of the dual-process theory. As Tinghög and others have remarked, was it introduced to generate testable predictions? Perhaps not; it appears researchers regard the dual-process theory more as a theoretical framework rather than a predictive model. Despite the level of empirical support, the theory offers a valuable starting point, enabling researchers to formulate more precise questions and develop more sophisticated theories.

Nevertheless, the dual-process theory may still represent a useful heuristic, given humanity's tendency to categorize the world in binaries: good versus bad, black versus white, fast versus slow, self versus others. With this in mind, does the dual-process theory enhance our understanding of human thought? Perhaps answering this question itself necessitates the engagement of System 2.[:]

Thinking: Beyond Fast and Slow

[:en]Thinking: Beyond Fast and Slow[:]

&nbsp;


	We all seem to have experienced this: when feeling down and dejected, a comforting pat on the back from a friend helps release our emotions; on a stormy night, hugging a stuffed toy allows us to sleep peacefully; and during a first date with a lover, the subtle intertwining of hands sparks feelings of affection&#8230;


	Touch is the earliest developed sense in the human body and is our most direct means of interacting with the world and others. It is not merely a sensory experience; it also triggers chemical reactions in the brain, releasing hormones and neurotransmitters that convey emotions, bringing us a sense of satisfaction and happiness. Touch plays an indispensable role in our lives.


	&nbsp;


	1. The Benefits of Touch


	Numerous studies have shown that touch benefits both mental and physical health. Touch can provide various physiological health benefits, including pain relief, blood pressure regulation, and cortisol reduction. Additionally, it can improve mental health by alleviating depression, reducing anxiety, and increasing positive emotions. However, because the type, pattern, and recipient of touch vary greatly, researchers have been exploring which kinds of touch are most effective for health.


	In a systematic review and meta-analysis published this year in Nature Human Behavior [1], the authors analyzed 212 studies on touch, comparing the effects of different types, frequencies, and durations of touch, as well as the health status and age of the recipients. The analysis revealed that, overall, touch is beneficial for health, but many factors influence its effectiveness. These findings also provide insights for future research on touch and robotics.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1937" height="271" src="https://admin.next-question.com/wp-content/uploads/2024/05/封面.png" width="573" srcset="https://admin.next-question.com/wp-content/uploads/2024/05/封面.png 573w, https://admin.next-question.com/wp-content/uploads/2024/05/封面-300x142.png 300w" sizes="(max-width: 573px) 100vw, 573px" />


	▷&nbsp;Packheiser, J., Hartmann, H., Fredriksen, K. et al. A systematic review and multivariate meta-analysis of the physical and mental health benefits of touch interventions. Nat Hum Behav (2024). https://doi.org/10.1038/s41562-024-01841-8


	&nbsp;


	It is important to note that the touch in these studies was conducted with the knowledge and consent of both parties. Ignoring or violating someone&#39;s consent for touch can have negative effects. Although there is limited research on non-consensual touch due to ethical reasons, merely imagining non-consensual touch can evoke feelings of dirtiness and a desire to cleanse oneself, accompanied by negative emotions such as depression, anxiety, anger, and shame [2]. This phenomenon is known as &quot;mental contamination.&quot; Therefore, consent is crucial in both research and practical applications of touch&mdash;do not touch others without their consent, no matter how much you desire it.


	&nbsp;


	2. How to Enhance Effectiveness?


	Research on adults indicates that physical touch significantly affects both physiological and mental health, with the most notable benefits being pain reduction and alleviation of depression and anxiety. Comparing different conditions of touch reveals more detailed insights.


	The location of touch influences the health benefits it provides. Touching the head, compared to the arms and torso, offers greater health benefits. Therefore, head touches, such as scalp or facial massages, may be particularly effective in improving health. There is no significant difference between massage therapy and other forms of touch, suggesting that various types of touch offer similar benefits. Increasing the frequency of touch sessions can bring more benefits, especially in alleviating anxiety, depression, and pain. However, extending the duration of each touch session does not enhance the benefits and may even diminish the regulation of cortisol and heart rate.


	Contextual factors such as the recipient&rsquo;s gender, cultural background, internal state (e.g., stress level), and familiarity with the toucher significantly influence the perception of affective touch. To maximize the pleasure and health benefits of touch, these factors must be considered. For example, studies have found that touch more effectively reduces cortisol in women and that hugs better alleviate physiological stress in women than in men. Notably, affective touch is not limited to romantic partners and can occur between friends, acquaintances, or even strangers. Overall, women are more receptive to touch from people with varying levels of familiarity, a gender difference particularly evident in studies on same-sex touch [3].


	Recent research also reveals that sensory pleasantness is influenced by gender and is related to the familiarity of the person providing the touch. However, since most studies on adult touch involve a majority of female participants, more research including male and non-binary participants is needed for more accurate conclusions. Additionally, older adults benefit more from touch in terms of improving positive emotions and lowering systolic blood pressure. This might be because, as age increases, baseline blood pressure also rises, providing more room for touch to have a regulatory effect.


	Although most studies focus on clinical populations, touch also benefits healthy individuals. However, clinical populations tend to gain more mental health improvements from touch, likely due to a higher need for contact. Loneliness often accompanies chronic illness and is associated with feelings of depression and anxiety. Touch can help mitigate these negative conditions.


	During the pandemic, social distancing measures were widely implemented to reduce COVID-19 transmission, inevitably limiting social interactions and touch. In the absence of touch, people might experience touch starvation, also known as skin hunger [4]. Studies have found that skin hunger is linked to stress, anxiety, and depression and may lead to aggressive behavior, self-harm, and eating disorders [5].


	The pandemic&rsquo;s restrictions on touch also prompt us to consider alternatives when human touch is not possible. Studies have found that body contact provided by objects (such as weighted blankets) or robots can also be beneficial. Physiologically, these benefits are comparable to those brought by human touch, but the mental health benefits are relatively less pronounced. Moreover, skin-to-skin contact has a better impact on mental health compared to other types of contact.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-medium wp-image-1936" height="300" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3-300x300.png" width="300" srcset="https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3-300x300.png 300w, https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3-150x150.png 150w, https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3-768x768.png 768w, https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3.png 983w" sizes="(max-width: 300px) 100vw, 300px" />


	▷&nbsp;Source: Andy Westface


	&nbsp;


	Because touch involves interpersonal relationships that contribute to positive health effects, research on using robots to provide touch has not only focused on how robots can assume social roles but also on how to improve their relationship with the touch recipient.


	&nbsp;


	Currently, common assistive robots are primarily divided into two categories: rehabilitation robots and socially assistive robots. Rehabilitation robots (e.g., intelligent wheelchairs, exoskeletons) primarily assist human rehabilitation through collaboration, while socially assistive robots cooperate directly with humans in touch-like interactions, performing tasks such as delivering items, cleaning, feeding, and providing companionship. These interactions offer psychological benefits and achieve mental health effects&nbsp;[6]. Additionally, researchers have developed synthetic skin in recent years to further enhance these mental health benefits by simulating skin contact.


	&nbsp;


	Findings from neonatal research differ from those in adults. Neonatal touch research often focuses more on physiological health, frequently studying the benefits of kangaroo care. Kangaroo care involves placing the newborn upright against the parent&rsquo;s chest, using the parent&rsquo;s body heat instead of an incubator to provide warmth and security. However, studies have found no significant difference between kangaroo care and other types of contact regarding health benefits for newborns, allowing for flexibility in choosing different types of touch based on the situation.


	&nbsp;


	Unlike in adults, where the familiarity of the toucher does not affect health outcomes, newborns benefit much more from being touched by their parents (often the mother in studies) than by strangers. Previous research has also indicated that early skin-to-skin contact and exposure to the mother&rsquo;s scent are crucial for newborns to adapt to their new environment. These findings highlight the irreplaceable role of parental care during this period. Additionally, the familiarity of the location where the touch occurs does not influence the effect of touch. Moreover, the frequency of touch does not apply to newborns as it does to adults. In terms of various physiological health aspects, touch is most effective in promoting weight gain in newborns.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-medium wp-image-1935" height="265" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture3-2-300x265.png" width="300" srcset="https://admin.next-question.com/wp-content/uploads/2024/05/Picture3-2-300x265.png 300w, https://admin.next-question.com/wp-content/uploads/2024/05/Picture3-2-768x679.png 768w, https://admin.next-question.com/wp-content/uploads/2024/05/Picture3-2.png 925w" sizes="(max-width: 300px) 100vw, 300px" />


	▷&nbsp;Touch benefits both adults&#39; and newborns&#39; mental and physical health. Source: Julian Packheiser


	&nbsp;


	3. Not Just for Humans!


	&nbsp;


	Animal welfare has become a significant social concern in recent years. On one hand, pets have become integral family members, playing important roles as companions in many people&rsquo;s lives. On the other hand, there is a growing awareness that even farm animals and laboratory mice should not be treated merely as tools; we should also strive to improve their well-being as much as possible within their limited lifespans.


	&nbsp;


	Consequently, some experiments have examined the effects of touch on animals. Most studies have been conducted on rodents, including rats, mice, monkeys, cats, sheep, and coral fish. The primary forms of touch studied are stroking (lightly touching with the hand) and tickling (touching a body part to induce spontaneous movements or laughter).


	&nbsp;


	Research has shown that touch can have positive physiological and quasi-psychological effects, indicating that human touch can also benefit animal health. For example, a 2014 study divided 139 anxious cats newly admitted to an animal shelter into a stroking group and a control group. The stroking group received gentle stroking (lightly touching the head and neck area while speaking softly) four times daily for ten days, each session lasting ten minutes [7]. The study found that the cats that were gently stroked daily exhibited fewer negative emotions (anxiety or irritation) compared to the control group and secreted more immunoglobulin A. Additionally, the incidence of upper respiratory diseases in the control group was 2.4 times higher than in the stroking group.


	&nbsp;


	Another interesting study used a machine to simulate the tactile stimulation provided by cleaner wrasses (Labroides dimidiatus) to striped surgeonfish (Ctenochaetus striatus) for removing ectoparasites&nbsp;[8]. The cortisol levels of the surgeonfish significantly decreased when they received tactile stimulation from the machine compared to when they were in contact with a stationary machine. The results suggest that even without social interaction, mere touch can yield positive health effects. These findings can help improve animal welfare in laboratory settings, animal husbandry, and pet relationships.


	&nbsp;


	Correspondingly, human-animal interaction is a crucial element of animal-assisted therapy, widely applied in treating conditions such as cognitive impairments, depression, and post-traumatic stress disorder. Studies have shown its potential for development in these areas [6,9].


	&nbsp;


	4. Conclusion


	&nbsp;


	In summary, touch is an innate ability and necessity for us. Whether it&rsquo;s a loving hug, a friendly handshake, or a parent&rsquo;s care for their child, we all have many warm memories associated with touch. However, in a world of rapid technological advancement, hectic work schedules, and sudden pandemics, we inevitably lose some opportunities for touch and, unknowingly, some of the positive emotions and sensations that accompany it. Hopefully, we can all utilize and enjoy touch, bringing comfort, relaxation, and happiness to ourselves and those around us during moments of fatigue and overwhelming stress.


	&nbsp;

[:en]
	&nbsp;



	We all seem to have experienced this: when feeling down and dejected, a comforting pat on the back from a friend helps release our emotions; on a stormy night, hugging a stuffed toy allows us to sleep peacefully; and during a first date with a lover, the subtle intertwining of hands sparks feelings of affection...



	Touch is the earliest developed sense in the human body and is our most direct means of interacting with the world and others. It is not merely a sensory experience; it also triggers chemical reactions in the brain, releasing hormones and neurotransmitters that convey emotions, bringing us a sense of satisfaction and happiness. Touch plays an indispensable role in our lives.



	&nbsp;



	1. The Benefits of Touch



	Numerous studies have shown that touch benefits both mental and physical health. Touch can provide various physiological health benefits, including pain relief, blood pressure regulation, and cortisol reduction. Additionally, it can improve mental health by alleviating depression, reducing anxiety, and increasing positive emotions. However, because the type, pattern, and recipient of touch vary greatly, researchers have been exploring which kinds of touch are most effective for health.



	In a systematic review and meta-analysis published this year in Nature Human Behavior [1], the authors analyzed 212 studies on touch, comparing the effects of different types, frequencies, and durations of touch, as well as the health status and age of the recipients. The analysis revealed that, overall, touch is beneficial for health, but many factors influence its effectiveness. These findings also provide insights for future research on touch and robotics.



	<img alt="" class="aligncenter size-full wp-image-1937" height="271" src="https://admin.next-question.com/wp-content/uploads/2024/05/封面.png" width="573" />



	▷&nbsp;Packheiser, J., Hartmann, H., Fredriksen, K. et al. A systematic review and multivariate meta-analysis of the physical and mental health benefits of touch interventions. Nat Hum Behav (2024). https://doi.org/10.1038/s41562-024-01841-8



	&nbsp;



	It is important to note that the touch in these studies was conducted with the knowledge and consent of both parties. Ignoring or violating someone&#39;s consent for touch can have negative effects. Although there is limited research on non-consensual touch due to ethical reasons, merely imagining non-consensual touch can evoke feelings of dirtiness and a desire to cleanse oneself, accompanied by negative emotions such as depression, anxiety, anger, and shame [2]. This phenomenon is known as &quot;mental contamination.&quot; Therefore, consent is crucial in both research and practical applications of touch&mdash;do not touch others without their consent, no matter how much you desire it.



	&nbsp;



	2. How to Enhance Effectiveness?



	Research on adults indicates that physical touch significantly affects both physiological and mental health, with the most notable benefits being pain reduction and alleviation of depression and anxiety. Comparing different conditions of touch reveals more detailed insights.



	The location of touch influences the health benefits it provides. Touching the head, compared to the arms and torso, offers greater health benefits. Therefore, head touches, such as scalp or facial massages, may be particularly effective in improving health. There is no significant difference between massage therapy and other forms of touch, suggesting that various types of touch offer similar benefits. Increasing the frequency of touch sessions can bring more benefits, especially in alleviating anxiety, depression, and pain. However, extending the duration of each touch session does not enhance the benefits and may even diminish the regulation of cortisol and heart rate.



	Contextual factors such as the recipient&rsquo;s gender, cultural background, internal state (e.g., stress level), and familiarity with the toucher significantly influence the perception of affective touch. To maximize the pleasure and health benefits of touch, these factors must be considered. For example, studies have found that touch more effectively reduces cortisol in women and that hugs better alleviate physiological stress in women than in men. Notably, affective touch is not limited to romantic partners and can occur between friends, acquaintances, or even strangers. Overall, women are more receptive to touch from people with varying levels of familiarity, a gender difference particularly evident in studies on same-sex touch [3].



	Recent research also reveals that sensory pleasantness is influenced by gender and is related to the familiarity of the person providing the touch. However, since most studies on adult touch involve a majority of female participants, more research including male and non-binary participants is needed for more accurate conclusions. Additionally, older adults benefit more from touch in terms of improving positive emotions and lowering systolic blood pressure. This might be because, as age increases, baseline blood pressure also rises, providing more room for touch to have a regulatory effect.



	Although most studies focus on clinical populations, touch also benefits healthy individuals. However, clinical populations tend to gain more mental health improvements from touch, likely due to a higher need for contact. Loneliness often accompanies chronic illness and is associated with feelings of depression and anxiety. Touch can help mitigate these negative conditions.



	During the pandemic, social distancing measures were widely implemented to reduce COVID-19 transmission, inevitably limiting social interactions and touch. In the absence of touch, people might experience touch starvation, also known as skin hunger [4]. Studies have found that skin hunger is linked to stress, anxiety, and depression and may lead to aggressive behavior, self-harm, and eating disorders [5].



	The pandemic&rsquo;s restrictions on touch also prompt us to consider alternatives when human touch is not possible. Studies have found that body contact provided by objects (such as weighted blankets) or robots can also be beneficial. Physiologically, these benefits are comparable to those brought by human touch, but the mental health benefits are relatively less pronounced. Moreover, skin-to-skin contact has a better impact on mental health compared to other types of contact.



	<img alt="" class="aligncenter size-medium wp-image-1936" height="300" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture2-3-300x300.png" width="300" />



	▷&nbsp;Source: Andy Westface



	&nbsp;



	Because touch involves interpersonal relationships that contribute to positive health effects, research on using robots to provide touch has not only focused on how robots can assume social roles but also on how to improve their relationship with the touch recipient.



	&nbsp;



	Currently, common assistive robots are primarily divided into two categories: rehabilitation robots and socially assistive robots. Rehabilitation robots (e.g., intelligent wheelchairs, exoskeletons) primarily assist human rehabilitation through collaboration, while socially assistive robots cooperate directly with humans in touch-like interactions, performing tasks such as delivering items, cleaning, feeding, and providing companionship. These interactions offer psychological benefits and achieve mental health effects&nbsp;[6]. Additionally, researchers have developed synthetic skin in recent years to further enhance these mental health benefits by simulating skin contact.



	&nbsp;



	Findings from neonatal research differ from those in adults. Neonatal touch research often focuses more on physiological health, frequently studying the benefits of kangaroo care. Kangaroo care involves placing the newborn upright against the parent&rsquo;s chest, using the parent&rsquo;s body heat instead of an incubator to provide warmth and security. However, studies have found no significant difference between kangaroo care and other types of contact regarding health benefits for newborns, allowing for flexibility in choosing different types of touch based on the situation.



	&nbsp;



	Unlike in adults, where the familiarity of the toucher does not affect health outcomes, newborns benefit much more from being touched by their parents (often the mother in studies) than by strangers. Previous research has also indicated that early skin-to-skin contact and exposure to the mother&rsquo;s scent are crucial for newborns to adapt to their new environment. These findings highlight the irreplaceable role of parental care during this period. Additionally, the familiarity of the location where the touch occurs does not influence the effect of touch. Moreover, the frequency of touch does not apply to newborns as it does to adults. In terms of various physiological health aspects, touch is most effective in promoting weight gain in newborns.



	<img alt="" class="aligncenter size-medium wp-image-1935" height="265" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture3-2-300x265.png" width="300" />



	▷&nbsp;Touch benefits both adults&#39; and newborns&#39; mental and physical health. Source: Julian Packheiser



	&nbsp;



	3. Not Just for Humans!



	&nbsp;



	Animal welfare has become a significant social concern in recent years. On one hand, pets have become integral family members, playing important roles as companions in many people&rsquo;s lives. On the other hand, there is a growing awareness that even farm animals and laboratory mice should not be treated merely as tools; we should also strive to improve their well-being as much as possible within their limited lifespans.



	&nbsp;



	Consequently, some experiments have examined the effects of touch on animals. Most studies have been conducted on rodents, including rats, mice, monkeys, cats, sheep, and coral fish. The primary forms of touch studied are stroking (lightly touching with the hand) and tickling (touching a body part to induce spontaneous movements or laughter).



	&nbsp;



	Research has shown that touch can have positive physiological and quasi-psychological effects, indicating that human touch can also benefit animal health. For example, a 2014 study divided 139 anxious cats newly admitted to an animal shelter into a stroking group and a control group. The stroking group received gentle stroking (lightly touching the head and neck area while speaking softly) four times daily for ten days, each session lasting ten minutes [7]. The study found that the cats that were gently stroked daily exhibited fewer negative emotions (anxiety or irritation) compared to the control group and secreted more immunoglobulin A. Additionally, the incidence of upper respiratory diseases in the control group was 2.4 times higher than in the stroking group.



	&nbsp;



	Another interesting study used a machine to simulate the tactile stimulation provided by cleaner wrasses (Labroides dimidiatus) to striped surgeonfish (Ctenochaetus striatus) for removing ectoparasites&nbsp;[8]. The cortisol levels of the surgeonfish significantly decreased when they received tactile stimulation from the machine compared to when they were in contact with a stationary machine. The results suggest that even without social interaction, mere touch can yield positive health effects. These findings can help improve animal welfare in laboratory settings, animal husbandry, and pet relationships.



	&nbsp;



	Correspondingly, human-animal interaction is a crucial element of animal-assisted therapy, widely applied in treating conditions such as cognitive impairments, depression, and post-traumatic stress disorder. Studies have shown its potential for development in these areas [6,9].



	&nbsp;



	4. Conclusion



	&nbsp;



	In summary, touch is an innate ability and necessity for us. Whether it&rsquo;s a loving hug, a friendly handshake, or a parent&rsquo;s care for their child, we all have many warm memories associated with touch. However, in a world of rapid technological advancement, hectic work schedules, and sudden pandemics, we inevitably lose some opportunities for touch and, unknowingly, some of the positive emotions and sensations that accompany it. Hopefully, we can all utilize and enjoy touch, bringing comfort, relaxation, and happiness to ourselves and those around us during moments of fatigue and overwhelming stress.



	&nbsp;

[:zh]


[:]

What Kinds of Touch are Most Beneficial for Mental and Physical Health?

[:en]What Kinds of Touch are Most Beneficial for Mental and Physical Health?[:]

In 2023, what has artificial intelligence brought to neuroscience?

[:en]In 2023, what has artificial intelligence brought to neuroscience?[:]

The hippocampus, the deity in charge of memory, I am your breaker of barriers.

[:en]The hippocampus, the deity in charge of memory, I am your breaker of barriers.[:zh]海马体掌管记忆的神，我是你的破壁人[:]

Recently, researchers from Tsinghua University, The Ohio State University, and the University of California, Berkeley, designed AgentBench, a testing tool for evaluating the reasoning and decision-making abilities of LLMs in a multi-turn, open-ended, generation setting. The researchers conducted a comprehensive evaluation of 25 LLMs, including API-based commercial and open-source models. They found that the top commercial LLMs demonstrated strong capabilities in complex environments. The researchers also said that AgentBench, a multidimensional evolving benchmark that currently consists of 8 distinct environments, will provide more in-depth systematic assessment of LLMs in the future.
&nbsp;
<h3>What are the settings assessed by the AgentBench?</h3>
AgentBench encompasses 8 distinct environments, 5 of which are created in this paper for the first time, namely operating system (OS), database (DB), knowledge graph (KG), digital card game (DCG), and lateral thinking puzzles (LTP). The remaining 3 are recompiled from published datasets, which are house-holding, web shopping, and web browsing.
All datasets are meticulously designed and reformulated to simulate interactive environments where text-only LLMs can operate as autonomous agents. Furthermore, they systematically evaluate an LLM’s core abilities, including following instructions, coding, knowledge acquisition, and logical reasoning. Compared to previous agent evaluation benchmarks, AgentBench concentrates on the practical use-oriented evaluation of LLMs via Chain-of-Thought (CoT) prompting.
&nbsp;
<h3>1. Operating System (OS)</h3>
In this dataset, the researchers aim to evaluate LLMs in genuine OS’ interactive bash environments on human questions with deterministic answers or series of operations for practical goals.
&nbsp;
<h3>2. Database (DB)</h3>
In AgentBench, the researchers evaluated LLMs on authentic SQL interfaces and databases as is in the real world. They have also measured the success rate of agents in completing instructions. Overall success rate is the macro average of the rate of three basic types of DB operations: select, insert, or update.
&nbsp;
<h3>3. Knowledge graph (KG)</h3>
Primarily, an intelligent agent needs to understand natural language, with all its intricacies and subtleties. It should also have the capacity to break down complex tasks into simpler, more manageable components. Operating in such environments, which are only partially observable, requires the agent to make decisions with incomplete information and manage inherent uncertainties. This heightens the need for the agent to demonstrate flexibility and adaptability in its decision making. Given the challenges outlined above, a KG can act as a representative testing ground to assess the decision-making abilities of AI agents in complex real-world environments.
&nbsp;
<h3>4. Digital card game (DCG)</h3>
Games, especially those requires strategies and planning, could serve as simulated environments for intelligent agent development. Digital card game is an ideal option for text-only LLM evaluation. It usually involves abundant text descriptions for cards, turn-based competition, and thoughtful playing strategies to win, testing a model’s understanding of game rules, operating logic, and abilities to form strategic decisions based on current conditions and past experiences in the game.
&nbsp;
<h3>5. Lateral thinking puzzles (LTP)</h3>
Lateral thinking puzzles, or namely situation puzzles, is a popular group-playing game around the world. Its name derives from the psychological term “lateral thinking”, which refers to the ability of deducing facts from unconventional perspectives and facilitating the exploration of new ideas. Through this assessment, the researchers aim to gain insights into the depth and agility of LLMs’ lateral reasoning abilities.
&nbsp;
<h3>6. House-Holding</h3>
In AgentBench, the researchers assess the model’s capability in accomplishing tasks in physical house-holding environments on the classical ALFWorld. The ALFWorld benchmark comprises of textual environments designed to mimic household scenarios, providing an interactive environment where an agent can perform decision-making tasks through text-based interfaces.
&nbsp;
<h3>7. Web shopping</h3>
Researchers used Webshop, a simulated online shopping environment, to evaluate language agents. The environment displays the text observation of the webpage and available actions to agents. Agent may freely explore the website and browse through items with clickable buttons just as in the real world.
&nbsp;
<h3>8. Web browsing</h3>
Mind2Web is a very recently released general benchmark for developing and assessing web agents capable of executing intricate tasks across various website domains, based on high-level user instructions. It designs feasible actions for website interactions, such as clicking, selecting, and typing, thereby facilitating a comprehensive evaluation of LLMs as web agents.
&nbsp;
<h3>Main results</h3>
On this challenging benchmark, surprisingly the researchers discover that some top LLMs have equipped strong capabilities in dealing with real-world environment interaction. For other API-based LLMs, despite their relatively poorer performance, regardless of tasks most of them can solve at least a few percent of problems. For open-sourced LLMs, however, it is common for them to fail to solve any problems in some challenging tasks.
This research unveiled that while top LLMs are becoming capable of tackling complex real-world missions, for open-sourced competitiors there are still a long way to go.It’s expected that AgentBench will serve as a cornerstone for later study to develop better and more applicable intelligent LLM agents.
<a href="https://arxiv.org/abs/2308.03688" target="_blank" rel="noopener">Read the paper on arXiv</a>

Scientists from Tsinghua University and UC Berkley reviewed 25 LLMs to identify their reasoning and decision-making abilities in an open-ended generation setting

[:en]Scientists from Tsinghua University and UC Berkley reviewed 25 LLMs to identify their reasoning and decision-making abilities in an open-ended generation setting[:zh]大模型能代替你刷知乎、打炉石传说、解谜“海龟汤”吗？清华、伯克利的科学家一口气测评了这25个LLM[:]

Because the Mountain is There - A Decade of Human Brain Projects

[:en]Because the Mountain is There - A Decade of Human Brain Projects[:zh]因为山就在那里——人类脑计划的十年[:]

In 2013, the National Institutes of Health (NIH) launched the Brain Research Through Advancing Innovative Neurotechnologies Initiative (BRAIN Initiative). In 2017, with the rapid development of single-cell sequencing and imaging technologies, the BICCN (BRAIN Initiative Cell Census Network) program was launched. It is a consortium of research teams located across the United States and Europe, aiming to use cutting-edge technologies to identify and classify cell types in the mouse, non-human primate, and human brains, as well as to develop new genetic tools for specific cell types.
&nbsp;
On October 13, 2023, scientists from the BICCN program published 21 papers in Science and its affiliated journals Science Advances and Science Translational Medicine, which released the largest and most comprehensive brain cell mapping of human and non-human primates, revealing more than 3,000 brain cell types, with many of these discoveries being the “first of their kind”.
&nbsp;
<h3>Single-cell mapping of adult human and non-human primates</h3>
 
The first study was led by neuroscientists Sten Linnarsson from the Karolinska Institute of Sweden and Kimberly Siletti from the University Medical Center Utrecht. They collected more than three million single cells from 106 sites in the forebrain, midbrain, and hindbrain of three deceased individuals for RNA sequencing (snRNA-seq), which provided insights into the diversity and subtypes of brain cells and laid a solid foundation for mapping the whole brain.
&nbsp;
By analyzing the sequencing data, they identified 461 broad categories of brain cells, including 3,313 subtypes, of which about 80% were neurons and the rest were different kinds of glial cells. Notably, the researchers found that the brainstem is the brain region with the greatest diversity of neurons and that some of these cells control innate behaviors such as pain reflexes, fear, aggression, and sexual behaviors, etc.
&nbsp;
Nikolas Jorstad from the Allen Institute for Brain Science in Seattle worked with his fellow co-workers to analyze eight regions of the cerebral cortex of adult humans and showed that most of the regions contain the same major cell types, with differences mainly in the proportions of each type and in the cortex where non-neuronal cells are located.
&nbsp;
Overall, these data support an evolutionary strategy of regional functional diversification that does not rely heavily on the generation of new cell types but rather utilizes subtle changes within cell types and in their relative distributions to construct distinct circuits.
&nbsp;
In addition to regional variability, there are individual differences in gene expression in brain cells. One of the studies reached this conclusion after comparing brain tissue from a large number of individuals using single-cell techniques for the first time. The results show that despite highly consistent cellular composition between individuals, there is a large amount of variation that can reflect individual characteristics, disease conditions, and genetic regulatory variants.
&nbsp;
The team of molecular biologist Joseph Ecker from the Salk Institute for Biological Studies in the U.S. chose to analyze the brain from an epigenetic perspective. In 2020, a team of researchers led by Professor Ecker analyzed more than 160 types of cells in the brains of mice, based on methylation markers on their DNAs, a methodology that can parse out the “on/off” states of genes. Ultimately, the team constructed a single-cell DNA methylation and three-dimensional genomic structural map of the human brain, elucidating the cellular specificity of various parts of the brain as well as their diverse epigenetic architectures.
&nbsp;
<h3>A comparative study of adult human and non-human primate single-cell mappings</h3>
&nbsp;
To clarify which features of brain cells are unique to humans versus non-human primates, Jorstad and other scientists performed single-cell transcriptional analysis of the middle temporal gyrus (a region critical for language comprehension) in adult humans, chimpanzees, gorillas, macaques, and common marmosets. They found that although primates largely share the same conserved cell types, they exhibit dramatic differences in cell proportions. Notably, although glial cells are generally less diverse than neurons within species, microglia, astrocytes, and oligodendrocytes showed greater transcriptional differences between species compared to neurons. These results suggest that specific properties of the adult human cortex may derive from relatively few cellular and molecular changes.
&nbsp;
Brian Lee et al. at the Allen Institute for Brain Science demonstrated that multimodal Patch-seq data are critical for refining the taxonomy of transcriptomic cell types using rapid viral gene labeling and patch clamp electrophysiology coupled with RNA sequencing (Patch-seq) to analyze neurons in neurosurgically resected living tissues.
&nbsp;
Thomas Chartrand et al, also from the Allen Institute for Brain Science, used single-cell transcriptomics to define neuronal cell types in cerebral cortex L1 of humans and mice and quantitatively identified homologous subspecies across species. The cross-species differences they observed suggest that humans and mice differ in their regulation of higher-order inputs to neocortical circuits.
&nbsp;
<h3>Modeling analysis of specific cell types of humans</h3>
&nbsp;
Tens of thousands of sequence variants in the human genome have been associated with the causes of neuropsychiatric disorders. To clarify their correlations, Ren Bing&#8217;s research team at the University of California, San Diego, conducted a comprehensive analysis of chromatin accessibility in the human brain at the single-cell level. Their results show that nearly one-third of transcriptional regulatory elements exhibit conservation and chromatin accessibility in mouse brain cells, emphasizing their functional importance.
&nbsp;
Single-cell developmental map of human and nonhuman primates
To track the developmental trajectories of the cerebral cortex, the research team of Arnold R. Kriegstein at the University of California, San Francisco, collected a large amount of small nucleus RNA sequencing (snRNA-seq) data from human cortical samples at different stages of development and performed single-cell trajectory analysis on single-cell data generated from human cortical samples at prenatal and postnatal developmental stages.
&nbsp;
This study elucidates the molecular changes behind human cortical lineage development. By integrating single-cell transcriptomic expression and chromatin accessibility mapping, they mapped a comprehensive cortical lineage covering prenatal and postnatal development, which identified key transcriptional networks and highlighted gender-specific developmental changes.
&nbsp;
They found that although neurons are the most diverse, both pre-astrocytes and oligodendrocyte precursor cells are also distinct across regions, with differences in their gene expression suggesting region-specific and cell-type-specific supporting functions. These findings emphasize the importance of early pattern events and provide a rich data resource for identifying therapeutic targets for human diseases affecting specific brain cell populations.
&nbsp;
Tomasz Nowakowski&#8217;s research team from the University of California, San Francisco, focused on analyzing features such as cell types and spatial organizations of the human thalamus as it develops. According to the analytical results, neurogenesis occurs in the thalamus during early gestation, giving rise to both glutamatergic and GABAergic neurons; by mid-gestation, glutamatergic neurons differentiate into two subtypes, while the number of GABAergic neurons in the thalamus increases dramatically and begins to be widely distributed. The results suggest that this population of GABAergic neurons may have contributed to human evolution.
&nbsp;
The breadth and depth of these BICCN studies demonstrate the potential of such large collaborative initiatives to generate fundamental knowledge of the human brain. This tremendous progress will provide invaluable resources for understanding how the human brain is formed and for research on human neurological disorders, as well as laying the groundwork for understanding the composition and function of the human brain and opening up a new era of inquiry into the etiology of human neurological disorders.
<a href="https://www.science.org/collections/brain-cell-census">Read related papers on Science</a>
&nbsp;

Milestones in Neuroscience! Comprehensive single-cell mapping of the human brain

[:en]Milestones in Neuroscience! Comprehensive single-cell mapping of the human brain[:zh]神经科学领域里程碑！全面构建人类大脑单细胞图谱[:]

Understanding how we learn during infancy not only helps us grasp the mechanisms of intellectual development but also inspires the design of training algorithms for machine learning. In a study published in Science in February 2024 [1], researchers examined how children associate language with observed objects by analyzing 61 hours of video recorded by cameras worn by infants. The researchers constructed a contrastive learning model to describe this learning process. Similarly, a 2023 NIPS paper [2] found that infants can spontaneously develop abstract representations from visual input through self-supervised learning. The next question is, how do infants achieve this step by step? This is precisely the question addressed by the new research introduced in this article.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1950" height="236" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture1.png" width="729" srcset="https://admin.next-question.com/wp-content/uploads/2024/06/Picture1.png 729w, https://admin.next-question.com/wp-content/uploads/2024/06/Picture1-300x97.png 300w" sizes="(max-width: 729px) 100vw, 729px" />


	▷&nbsp;Anderson, Erin M., et al. &quot;An edge-simplicity bias in the visual input to young infants.&quot; Science Advances 10.19 (2024): eadj8571.


	&nbsp;


	A paper published on May 10, 2024, in Science Advances [3] employed a similar experimental design, using cameras mounted on infants&#39; heads to directly observe and analyze the visual stimuli in their environments. By comparing the images observed by infants with those perceived by adults, the study found that infants experience unique visual inputs consisting of simple, high-contrast patterns and their edges. These high-contrast patterns are crucial for the development of visual modules in the brain.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1949" height="356" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture2.png" width="605" srcset="https://admin.next-question.com/wp-content/uploads/2024/06/Picture2.png 605w, https://admin.next-question.com/wp-content/uploads/2024/06/Picture2-300x177.png 300w" sizes="(max-width: 605px) 100vw, 605px" />


	▷&nbsp;Figure 1: Experimental design showing that infants, in early visual development, tend to observe patterns composed of a few high-contrast edges as in figure b. Source: Reference 3.


	&nbsp;


	The traditional assumption in cognitive science is that visual input is fundamentally the same for everyone, regardless of developmental stage. However, the new research, based on data recorded by head-mounted cameras on 10 infants aged 3-13 weeks (5 males) and a control group of 10 adults aged 31-70 years, found that visual input varies with development. The visual experiences of very young infants are specific to their age. They prefer simple, high-contrast scenes (Figure 2), such as broad black stripes and checkerboard patterns.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1948" height="529" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture3.png" width="605" srcset="https://admin.next-question.com/wp-content/uploads/2024/06/Picture3.png 605w, https://admin.next-question.com/wp-content/uploads/2024/06/Picture3-300x262.png 300w" sizes="(max-width: 605px) 100vw, 605px" />


	▷&nbsp;Figure 2: Infants prefer visual input with high contrast and simple edges. Source: Reference 3.


	&nbsp;


	Researchers divided the images into four quadrants based on simplicity and contrast. They found that infants preferred patterns with simple boundaries and high contrast (Figure 3). This provides insight into what attracts the attention of infants.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-1947" height="317" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture4.png" width="605" srcset="https://admin.next-question.com/wp-content/uploads/2024/06/Picture4.png 605w, https://admin.next-question.com/wp-content/uploads/2024/06/Picture4-300x157.png 300w" sizes="(max-width: 605px) 100vw, 605px" />


	▷&nbsp;Figure 3: Proportion of observation of images with different feature combinations by infants and adults. Source: Reference 3.


	&nbsp;


	Given that the V1 area of the visual cortex extracts local edges to help construct meaningful objects and scenes, could infants&#39; visual preferences guide the training of visual models? A 2023 NIPS paper [4]&nbsp;found that using images similar to those observed by infants during early development, rather than random patterns seen by adults, improved the performance of AI systems in visual recognition tasks. This study used images captured by cameras worn by infants, although researchers at the time were unaware of infants&#39; visual preferences. The new research suggests that a preference for simplicity and high contrast might be beneficial for training the V1 area during early visual development. The convolutional neural network architecture used in AI recognition, modeled after the human visual cortex, shows similar characteristics.


	The study subjects were children of faculty members from Indiana University Birmingham, living in an artificially constructed environment. This raises the question: could the study&#39;s conclusions be limited to infants in such environments, lacking cross-cultural universality? A 2023 study&nbsp;[5]&nbsp;provides a rebuttal by comparing head-mounted camera data from infants in a small, crowded fishing village in Sihai, India. With limited electricity, most daily activities occur outdoors. The results showed no statistical difference between the data observed by infants in Sihai and those in the West, with children from both regions favoring simple, high-contrast patterns.


	This research suggests that infants&#39; visual observation patterns could be used to identify and intervene early in the development of diseases such as cataracts, strabismus, refractive errors, and ptosis. These conditions disrupt visual development by interfering with input to the visual cortex, leading to abnormalities. In the future, infants could wear cameras, and algorithms could detect if they do not show a preference for simple and high-contrast images, allowing for low-cost identification of related diseases.


	While mammals like horses and sheep can run shortly after birth, human infants take about three months to hear and see, and another six months to control their posture and head slightly. Why do humans require such a long time for their nervous system to mature? The study offers a possible explanation: the visual system first trains the V1 area for edge recognition, followed by training the V2-V6 areas for more abstract representations. This slow, step-by-step optimization helps build a more intelligent and flexible visual system.


	Following this reasoning, young primates such as gorillas could be fitted with cameras during their critical period of visual development to see if they exhibit visual preferences similar to human infants. Although no such studies have been found, a 2020 study on various primates, including the smallest primate&nbsp;[6], the mouse lemur, found that the arrangement of visual processing units in the brain follows the same mathematical rules across primates of different sizes.


	Therefore, it is reasonable to infer that similar patterns might be observed in primates like gorillas. However, for animals like horses and sheep that can move immediately after birth, it may not be possible to observe a preference for simple and high-contrast patterns in their young. For animals like cats and dogs, which also require a development period (from opening their eyes to normal walking and hunting) but are not primates, it is difficult to predict.


	Cross-species comparative studies will reveal how the visual systems of different organisms balance the higher intelligence brought by slow development with the evolutionary advantages of innate abilities. This is a core topic for machine intelligence, extending beyond the visual system to the broader discussion of innate versus acquired traits.


	&nbsp;

[:en]
	Understanding how we learn during infancy not only helps us grasp the mechanisms of intellectual development but also inspires the design of training algorithms for machine learning. In a study published in Science in February 2024 [1], researchers examined how children associate language with observed objects by analyzing 61 hours of video recorded by cameras worn by infants. The researchers constructed a contrastive learning model to describe this learning process. Similarly, a 2023 NIPS paper [2] found that infants can spontaneously develop abstract representations from visual input through self-supervised learning. The next question is, how do infants achieve this step by step? This is precisely the question addressed by the new research introduced in this article.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-1950" height="236" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture1.png" width="729" />



	▷&nbsp;Anderson, Erin M., et al. &quot;An edge-simplicity bias in the visual input to young infants.&quot; Science Advances 10.19 (2024): eadj8571.



	&nbsp;



	A paper published on May 10, 2024, in Science Advances [3] employed a similar experimental design, using cameras mounted on infants&#39; heads to directly observe and analyze the visual stimuli in their environments. By comparing the images observed by infants with those perceived by adults, the study found that infants experience unique visual inputs consisting of simple, high-contrast patterns and their edges. These high-contrast patterns are crucial for the development of visual modules in the brain.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-1949" height="356" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture2.png" width="605" />



	▷&nbsp;Figure 1: Experimental design showing that infants, in early visual development, tend to observe patterns composed of a few high-contrast edges as in figure b. Source: Reference 3.



	&nbsp;



	The traditional assumption in cognitive science is that visual input is fundamentally the same for everyone, regardless of developmental stage. However, the new research, based on data recorded by head-mounted cameras on 10 infants aged 3-13 weeks (5 males) and a control group of 10 adults aged 31-70 years, found that visual input varies with development. The visual experiences of very young infants are specific to their age. They prefer simple, high-contrast scenes (Figure 2), such as broad black stripes and checkerboard patterns.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-1948" height="529" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture3.png" width="605" />



	▷&nbsp;Figure 2: Infants prefer visual input with high contrast and simple edges. Source: Reference 3.



	&nbsp;



	Researchers divided the images into four quadrants based on simplicity and contrast. They found that infants preferred patterns with simple boundaries and high contrast (Figure 3). This provides insight into what attracts the attention of infants.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-1947" height="317" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture4.png" width="605" />



	▷&nbsp;Figure 3: Proportion of observation of images with different feature combinations by infants and adults. Source: Reference 3.



	&nbsp;



	Given that the V1 area of the visual cortex extracts local edges to help construct meaningful objects and scenes, could infants&#39; visual preferences guide the training of visual models? A 2023 NIPS paper [4]&nbsp;found that using images similar to those observed by infants during early development, rather than random patterns seen by adults, improved the performance of AI systems in visual recognition tasks. This study used images captured by cameras worn by infants, although researchers at the time were unaware of infants&#39; visual preferences. The new research suggests that a preference for simplicity and high contrast might be beneficial for training the V1 area during early visual development. The convolutional neural network architecture used in AI recognition, modeled after the human visual cortex, shows similar characteristics.



	The study subjects were children of faculty members from Indiana University Birmingham, living in an artificially constructed environment. This raises the question: could the study&#39;s conclusions be limited to infants in such environments, lacking cross-cultural universality? A 2023 study&nbsp;[5]&nbsp;provides a rebuttal by comparing head-mounted camera data from infants in a small, crowded fishing village in Sihai, India. With limited electricity, most daily activities occur outdoors. The results showed no statistical difference between the data observed by infants in Sihai and those in the West, with children from both regions favoring simple, high-contrast patterns.



	This research suggests that infants&#39; visual observation patterns could be used to identify and intervene early in the development of diseases such as cataracts, strabismus, refractive errors, and ptosis. These conditions disrupt visual development by interfering with input to the visual cortex, leading to abnormalities. In the future, infants could wear cameras, and algorithms could detect if they do not show a preference for simple and high-contrast images, allowing for low-cost identification of related diseases.



	While mammals like horses and sheep can run shortly after birth, human infants take about three months to hear and see, and another six months to control their posture and head slightly. Why do humans require such a long time for their nervous system to mature? The study offers a possible explanation: the visual system first trains the V1 area for edge recognition, followed by training the V2-V6 areas for more abstract representations. This slow, step-by-step optimization helps build a more intelligent and flexible visual system.



	Following this reasoning, young primates such as gorillas could be fitted with cameras during their critical period of visual development to see if they exhibit visual preferences similar to human infants. Although no such studies have been found, a 2020 study on various primates, including the smallest primate&nbsp;[6], the mouse lemur, found that the arrangement of visual processing units in the brain follows the same mathematical rules across primates of different sizes.



	Therefore, it is reasonable to infer that similar patterns might be observed in primates like gorillas. However, for animals like horses and sheep that can move immediately after birth, it may not be possible to observe a preference for simple and high-contrast patterns in their young. For animals like cats and dogs, which also require a development period (from opening their eyes to normal walking and hunting) but are not primates, it is difficult to predict.



	Cross-species comparative studies will reveal how the visual systems of different organisms balance the higher intelligence brought by slow development with the evolutionary advantages of innate abilities. This is a core topic for machine intelligence, extending beyond the visual system to the broader discussion of innate versus acquired traits.



	&nbsp;

[:zh]


[:]

From Comparing the World: The Common Language of Infants and Artificial Intelligence

[:en]From Comparing the World: The Common Language of Infants and Artificial Intelligence[:]

It has taken decades for computer scientists to finally materialize automatic speech recognition AI models that are closer to the human level. Such purely engineered AI models have completely abandoned earlier modeling frameworks based on linguistic theory, and entirely adopted data-driven, end-to-end, large-scale pre-trained deep neural networks. So how much does such a model resemble the auditory pathways of the human brain?
&nbsp;
To address this issue, Professor Yuanning Li&#8217;s team at the School of Biomedical Engineering, ShanghaiTech University, in collaboration with Professor Edward Chang at the University of California, San Francisco, and Professor Jinsong Wu/Junfeng Lu&#8217;s team at Fudan University, integrated a variety of technological approaches such as self-supervised pre-trained in-depth speech modeling, high-density intracranial electroencephalograms, and single-neuron simulation modeling, and thoroughly investigated the computational and representational similarities between AI speech models and the auditory pathways of the human brain utilizing a cross-language (English and Chinese) control experiment paradigm.
&nbsp;
Researchers can extract features at different levels &#8211; from purely acoustic spectrogram features to phonetic vowels and consonants and articulations, to relative pitches containing contextual information, etc. &#8211; and then use sliding time windows to predict neural responses. If a particular type of feature accurately predicts neural activity in a particular region, it is usually assumed that the neural activity encoded in that region expresses that type of feature.
&nbsp;
Over the past decade or so, using intracranial electrophysiological recording experiments as well as neural coding models, researchers have identified several important features of neural coding, e.g., the activity of different neural populations in the secondary auditory cortex of the superior temporal gyrus encodes features ranging from envelopes of voice to specific vowel and consonant phonemes, and so on.
&nbsp;
In this study, in addition to the secondary auditory cortex, which is closely related to speech, using intracranial high-density EEG recording technology and high-precision single-neuron biophysical simulation modeling, the researchers obtained neural responses covering the entire auditory pathway from the auditory nerves to the brainstem to the auditory cortex.
&nbsp;
Given that AI models and brain auditory circuits are capable of receiving the same speech input and performing similar cognitive functions, are there computational and representational similarities between the two? This is the key question on which this study focuses.
&nbsp;
To address this problem, the researchers have constructed a new deep neural coding model. This is a purely data-driven model that extracts feature representations from deep neural networks pre-trained with speeches, applies these data-driven features to construct a new linear coding model, and performs correlation analysis with authentic brain auditory response signals to investigate the similarity between the intrinsic feature representations of deep neural networks and the activities of different neural populations within the brain auditory pathway.
&nbsp;
By comparing the performance of neural activity prediction at different nodes of the auditory pathway of the neural coding models built based on these models, it was found that there is indeed a great similarity between the hierarchical structure of the end-to-end speech pre-training network, and that of the auditory circuit.
&nbsp;
First, for the entire auditory pathway, the encoding prediction model based on deep neural network features is comprehensively better than the traditional linear feature model based on linguistic theory. This indicates that the entire auditory pathway has strong nonlinear features. Second, models with different levels of complexity correspond to different regions in the auditory pathway. In addition, it was found that for the same self-supervised speech model, its overall hierarchical structure corresponds to the AN-IC-STG hierarchical structure of the auditory pathway.
&nbsp;
Having established the representational similarities between the deep speech model and the auditory pathway, the researchers further explored the computational mechanisms driving these representational similarities, focusing on the best-performing HuBERT model.
&nbsp;
The results show that as the network deepens, the attentional weights aligned to long-range contextual structures become progressively larger. It is worth emphasizing that the HuBERT model used here is fully self-supervised, and the training process does not include any explicit information about the context structure as well as the speech content information. This result suggests that a self-supervised trained speech model can learn key context structure information related to language and semantics in natural speech.
&nbsp;
In the secondary auditory cortex of the superior temporal gyrus, which is closely related to speech processing, the higher the alignment of the self-attentional weights with the contextual structures in speech, the better the neural network&#8217;s prediction of brain activity, while conversely, in the primary auditory cortex as well as in these regions of the auditory nerve and brainstem, the greater the expression of localized transient information in the temporal domain, and the greater the resemblance between neural networks and brain signals.
&nbsp;
Last but not least, the study further analyzed whether the self-supervised model was able to learn more advanced contextual information. The results show that the self-supervised model is able to learn higher-level contextual information related to linguistic specificity and that this specificity information is significantly correlated with computation and representation in the phonetic cortex of the brain.
&nbsp;
This study provides a new biological perspective on unlocking the “black box” of deep neural networks, especially the self-attention model called Transformer.
&nbsp;
<a href="https://www.nature.com/articles/s41593-023-01468-4">Read the paper on Nature Neuroscience</a>

How similar are AI speech models to the auditory pathways of humans?

[:en]How similar are AI speech models to the auditory pathways of humans?[:zh]AI语音模型与人的听觉有多相似？上科大/UCSF/复旦联合团队解析深度语音模型与人脑听觉通路的表征与计算相似性[:]

&nbsp;


	&nbsp;


	Cognitive science traditionally categorizes human behaviors and those of algorithm-driven agents into two types: one is goal-driven, like an explorer with a map who knows where they are headed; the other follows habits, like a student who consistently takes the same route to school. It has been generally accepted among researchers that these two types of behavior are governed by distinct neural mechanisms. However, a recent study published in Nature Communications proposes that both types of behavior can be unified under the same theoretical framework: variational Bayesian theory. <img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2396" height="302" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture1.png" width="789" srcset="https://admin.next-question.com/wp-content/uploads/2024/09/Picture1.png 789w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture1-300x115.png 300w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture1-768x294.png 768w" sizes="(max-width: 789px) 100vw, 789px" /> ▷ Han, D., Doya, K., Li, D. et al. Synergizing habits and goals with variational Bayes. Nat Commun 15, 4461 (2024). https://doi.org/10.1038/s41467-024-48577-7


	&nbsp;

<h2 style="line-height: 1.5;">
	1. Question: The Same Domain, the Same Essence 
</h2>

	In scientific research, whether studying animals, humans, or machine learning algorithms, we often encounter the question: When faced with a new environment or challenge, should we rely on instincts and habits, or should we try to learn new methods? This question may seem to span various fields, but in reality, these fields all revolve around a central issue: How to strike the optimal balance between rapid adaptation and maintaining flexibility.


	For example, when animals face environmental changes, they may instinctively react by seeking shelter or searching for food&mdash;behaviors that may be innate or learned. Similarly, in decision-making, people sometimes rely on intuition (what psychologist Daniel Kahneman calls &quot;System 1&quot;), while other times they engage in more careful deliberation (&quot;System 2&quot;). In machine learning, some algorithms are &quot;model-free,&quot; meaning they learn from experience without predefined rules, while others are &quot;model-based,&quot; relying on explicit rules and models. <img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2397" height="359" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture2.png" width="691" srcset="https://admin.next-question.com/wp-content/uploads/2024/09/Picture2.png 691w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture2-300x156.png 300w" sizes="(max-width: 691px) 100vw, 691px" /> ▷ The characteristics of habitual behaviors (e.g., snacking while focused on work) and goal-directed behaviors (e.g., planning a diet meal).


	These seemingly different situations are, in fact, quite similar: whether in biological or artificial systems, the key challenge is to balance rapid adaptation with flexibility in handling new situations.


	For instance, bacteria quickly adapt to their environment through chemotaxis (the instinct to move toward nutrient-rich areas). When faced with an environment without pre-established rules, an intelligent agent may behave like a student immersed in solving problems, quickly reaching the goal. However, more complex organisms or algorithms may require more flexible strategies to tackle more sophisticated challenges.


	To explain the process of goal-directed learning, neuroscientists have proposed the theoretical framework of active inference. This framework posits that the brain is constantly trying to minimize uncertainty and unexpected outcomes when predicting the environment by directing the body to interact with it. The core concept of this theory, &quot;free energy,&quot; measures the difference between the probability of sensory inputs and the expected sensory inputs. The process of active inference is essentially the process of minimizing free energy.


	While &quot;active inference&quot; provides insight into goal-directed learning, it remains a hypothesis that has yet to be fully validated in the scientific community. There is still insufficient empirical evidence to support the neural mechanisms behind it. For instance, active inference explains goal-directed behaviors as a process of minimizing the gap between goals and reality. However, it falls short in explaining habit-based behaviors that do not require conscious intervention or rely on external feedback.


	Recent studies have attempted to demonstrate how these seemingly opposing behavioral modes, goal-directed and habit-driven, can work together within a unified theoretical framework, enabling organisms to efficiently and flexibly adapt to their environment.


	&nbsp;

<h2 style="line-height: 1.5;">
	2. Discovery: Predictive Coding and Complexity Reduction&mdash;The Brain&#39;s Ongoing Bayesian Inference 
</h2>

	To better understand this new framework, we can liken the brain to a chef constantly experimenting with new dishes. When a chef adjusts his menu, is he making the dishes more appealing to customers or simply cooking out of habit? In reality, he is doing both. On the one hand, he reduces the gap between his culinary creations and the customers&#39; tastes. On the other hand, he continuously updates and simplifies the predictive model of changing customer preferences&mdash;his expectations of their tastes.


	This process can be illustrated with a simple example. A limited menu with only a few dishes may lack the flexibility to satisfy all customers&#39; needs. In contrast, a complex menu that can be freely adjusted based on customer feedback might better accommodate different tastes, but it could become too complicated to manage, increasing costs or leading to inconsistent flavors.


	Scientists have described this process in mathematical terms, defining &quot;latent intentions&quot; to expand the concept of free energy. In this context, free energy includes not only the concept from active inference but also the agent&#39;s behavioral tendencies and predictions about observations. The agent&#39;s learning behavior can be seen as a continuous updating process (Markov chain) aimed at minimizing the value of Zt. This value comprises prediction error (the discrepancy with reality) and KL divergence (the complexity of the model).


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2398" height="124" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture3.png" width="605" srcset="https://admin.next-question.com/wp-content/uploads/2024/09/Picture3.png 605w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture3-300x61.png 300w" sizes="(max-width: 605px) 100vw, 605px" />


	British cognitive scientist Andy Clark pointed out that the brain is a powerful prediction machine, constantly forecasting upcoming sensory inputs and adjusting these predictions based on actual inputs. In this process, the prediction error corresponds to the first term in the aforementioned formula. The second term, KL divergence, measures the difference in probability density between predictions before and after an action, reflecting the complexity of the predictive model. In habit-driven learning, whether an action is taken or not does not affect the prediction, making this term zero, meaning no model exists. Thus, KL divergence, which represents model complexity, turns the binary distinction between model-free and model-based approaches into a continuous spectrum.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2399" height="317" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture4.png" width="605" srcset="https://admin.next-question.com/wp-content/uploads/2024/09/Picture4.png 605w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture4-300x157.png 300w" sizes="(max-width: 605px) 100vw, 605px" />


	▷ a) Schematic of integrating habits and goal frameworks. b) Structure of the framework during training. c) Structure of the framework during behavioral processes.


	In this framework, when faced with a new goal, early learning resembles a model-based system, similar to a chef who, upon opening his restaurant, tries to optimize his predictions of new customers&#39; tastes. Once this predictive model has been sufficiently refined through continuous training, it may shift toward a more habit-driven approach, where the chef continually perfects his signature dishes.


	&nbsp;

<h2 style="line-height: 1.5;">
	3. Significance: Enabling AI to Perform Zero-Shot Learning 
</h2>

	One significant advantage of human intelligence is the ability to solve various tasks in entirely new environments without relying on prior examples. For instance, when a painter is asked to depict a mythical creature like a qilin, he or she only needs to know that the qilin is a symbol of good fortune. However, an AI would require specific prompts, such as: &quot;Depict a qilin from ancient Chinese mythology, with a dragon&#39;s head, deer antlers, lion&#39;s eyes, tiger&#39;s back, bear&#39;s waist, snake&#39;s scales, horse&#39;s hooves, and an ox&#39;s tail, exuding an overall aura of majesty and sanctity, with gold and red as the primary colors, against a backdrop of swirling clouds, symbolizing good fortune, peace, and imperial power.&quot;


	The first painter to depict a qilin did so through zero-shot learning after acquiring sufficient experience in painting; however, zero-shot learning remains a challenging task for current AI systems. This is precisely the issue that this framework aims to address.


	When the environment changes, an agent built using the integrated framework proposed in this paper can spontaneously switch from habit-based model-free learning to model-based learning. This allows the agent to adapt to the new environmental conditions. In their experiments, researchers used a T-maze to test the adaptability of these agents. In this maze, the agent must decide which direction to take based on rewards on either side, learning strategies to maximize rewards.


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2400" height="190" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture5.png" width="500" srcset="https://admin.next-question.com/wp-content/uploads/2024/09/Picture5.png 500w, https://admin.next-question.com/wp-content/uploads/2024/09/Picture5-300x114.png 300w" sizes="(max-width: 500px) 100vw, 500px" /> ▷ T-maze and the three stages of the agent in response to environmental changes.


	In a habit-based system, the agent might continue following the previously rewarded path, even if the reward has changed. However, goal-directed agents face a different challenge. For example, if the initial reward on the left side of the maze is 100 times greater than on the right, the agent might need to try the left side a hundred times before updating the model and attempting the right side (depending on the specific agent model). This is undoubtedly an inefficient approach; in the real world, an organism displaying such behavior would likely be eliminated by natural selection. The framework proposed in this paper integrates goal-driven and habit-driven approaches, allowing the agent to balance flexibility and speed. Initially, it adapts to the environment by choosing the right side; when the environment changes (the left reward disappears), it readjusts to choose the left side.


	The simple T-maze experiment shows that the new framework aligns with Yann LeCun&#39;s concept of the world model. LeCun emphasizes that a world model has a dual role of planning for the future and estimating missing observations, and it should be an energy-based model. In goal-directed behavior, this framework uses the current state, goal, and intended actions as inputs, and outputs an energy value to describe their &quot;consistency.&quot; It can be said that the agent&#39;s decision-making in the T-maze builds upon and relies on the world model envisioned by LeCun.


	There is a long road ahead from the simple T-maze to highly complex large language models. However, based on the theoretical framework described in this paper, we can observe some important similarities. For example, in training language models, we typically predict based on existing vocabulary, similar to scenarios without specific goals set during the training phase. The flexibility of goal-directed planning in this framework comes from its ability to break down any future goal into a series of sequential steps, predicting only the next observation. This method minimizes the discrepancy between goal-directed intentions and prior distributions, thereby compressing the search space and making the search process more efficient. This approach is also suitable for large models.


	Additionally, based on the KL divergence term in the framework, we can understand the hierarchical structure in predictive coding, where a hierarchical information processing method is used to reduce model complexity. Predictive coding theory also suggests that the brain learns to recognize patterns by filtering out information that can be predicted through natural world patterns, thereby reducing unnecessary data. This information processing strategy echoes the information bottleneck theory, showing how cognitive processes can be optimized by using lower-dimensional representations.


	Finally, this theoretical framework not only enhances our understanding of healthy brain function but also provides new perspectives for understanding and treating neurological disorders. For example, patients with Parkinson&#39;s disease often struggle with goal-directed planning abilities and rely more on habitual behaviors. This may be due to high uncertainty within goal-directed intentions. Research on how medical interventions or deep brain stimulation (altering internal states) and sensory stimulation (altering brain input) can reduce this uncertainty may provide methods to improve motor control in Parkinson&#39;s patients.


	Moreover, research on autism spectrum disorder (ASD) can also benefit from this theoretical framework. Individuals with autism often exhibit repetitive behaviors, which may be related to an overemphasis on model complexity in predictive coding, affecting cognitive behavioral flexibility when adapting to changing environments. Introducing some randomness to increase behavioral diversity could be a potential intervention strategy.


	&nbsp;


	&nbsp;

[:en]
	&nbsp;



	&nbsp;



	Cognitive science traditionally categorizes human behaviors and those of algorithm-driven agents into two types: one is goal-driven, like an explorer with a map who knows where they are headed; the other follows habits, like a student who consistently takes the same route to school. It has been generally accepted among researchers that these two types of behavior are governed by distinct neural mechanisms. However, a recent study published in Nature Communications proposes that both types of behavior can be unified under the same theoretical framework: variational Bayesian theory. <img alt="" class="aligncenter size-full wp-image-2396" height="302" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture1.png" width="789" /> ▷ Han, D., Doya, K., Li, D. et al. Synergizing habits and goals with variational Bayes. Nat Commun 15, 4461 (2024). https://doi.org/10.1038/s41467-024-48577-7



	&nbsp;


<h2 style="line-height: 1.5;">
	1. Question: The Same Domain, the Same Essence
</h2>


	In scientific research, whether studying animals, humans, or machine learning algorithms, we often encounter the question: When faced with a new environment or challenge, should we rely on instincts and habits, or should we try to learn new methods? This question may seem to span various fields, but in reality, these fields all revolve around a central issue: How to strike the optimal balance between rapid adaptation and maintaining flexibility.



	For example, when animals face environmental changes, they may instinctively react by seeking shelter or searching for food&mdash;behaviors that may be innate or learned. Similarly, in decision-making, people sometimes rely on intuition (what psychologist Daniel Kahneman calls &quot;System 1&quot;), while other times they engage in more careful deliberation (&quot;System 2&quot;). In machine learning, some algorithms are &quot;model-free,&quot; meaning they learn from experience without predefined rules, while others are &quot;model-based,&quot; relying on explicit rules and models. <img alt="" class="aligncenter size-full wp-image-2397" height="359" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture2.png" width="691" /> ▷ The characteristics of habitual behaviors (e.g., snacking while focused on work) and goal-directed behaviors (e.g., planning a diet meal).



	These seemingly different situations are, in fact, quite similar: whether in biological or artificial systems, the key challenge is to balance rapid adaptation with flexibility in handling new situations.



	For instance, bacteria quickly adapt to their environment through chemotaxis (the instinct to move toward nutrient-rich areas). When faced with an environment without pre-established rules, an intelligent agent may behave like a student immersed in solving problems, quickly reaching the goal. However, more complex organisms or algorithms may require more flexible strategies to tackle more sophisticated challenges.



	To explain the process of goal-directed learning, neuroscientists have proposed the theoretical framework of active inference. This framework posits that the brain is constantly trying to minimize uncertainty and unexpected outcomes when predicting the environment by directing the body to interact with it. The core concept of this theory, &quot;free energy,&quot; measures the difference between the probability of sensory inputs and the expected sensory inputs. The process of active inference is essentially the process of minimizing free energy.



	While &quot;active inference&quot; provides insight into goal-directed learning, it remains a hypothesis that has yet to be fully validated in the scientific community. There is still insufficient empirical evidence to support the neural mechanisms behind it. For instance, active inference explains goal-directed behaviors as a process of minimizing the gap between goals and reality. However, it falls short in explaining habit-based behaviors that do not require conscious intervention or rely on external feedback.



	Recent studies have attempted to demonstrate how these seemingly opposing behavioral modes, goal-directed and habit-driven, can work together within a unified theoretical framework, enabling organisms to efficiently and flexibly adapt to their environment.



	&nbsp;


<h2 style="line-height: 1.5;">
	2. Discovery: Predictive Coding and Complexity Reduction&mdash;The Brain&#39;s Ongoing Bayesian Inference
</h2>


	To better understand this new framework, we can liken the brain to a chef constantly experimenting with new dishes. When a chef adjusts his menu, is he making the dishes more appealing to customers or simply cooking out of habit? In reality, he is doing both. On the one hand, he reduces the gap between his culinary creations and the customers&#39; tastes. On the other hand, he continuously updates and simplifies the predictive model of changing customer preferences&mdash;his expectations of their tastes.



	This process can be illustrated with a simple example. A limited menu with only a few dishes may lack the flexibility to satisfy all customers&#39; needs. In contrast, a complex menu that can be freely adjusted based on customer feedback might better accommodate different tastes, but it could become too complicated to manage, increasing costs or leading to inconsistent flavors.



	Scientists have described this process in mathematical terms, defining &quot;latent intentions&quot; to expand the concept of free energy. In this context, free energy includes not only the concept from active inference but also the agent&#39;s behavioral tendencies and predictions about observations. The agent&#39;s learning behavior can be seen as a continuous updating process (Markov chain) aimed at minimizing the value of Zt. This value comprises prediction error (the discrepancy with reality) and KL divergence (the complexity of the model).



	<img alt="" class="aligncenter size-full wp-image-2398" height="124" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture3.png" width="605" />



	British cognitive scientist Andy Clark pointed out that the brain is a powerful prediction machine, constantly forecasting upcoming sensory inputs and adjusting these predictions based on actual inputs. In this process, the prediction error corresponds to the first term in the aforementioned formula. The second term, KL divergence, measures the difference in probability density between predictions before and after an action, reflecting the complexity of the predictive model. In habit-driven learning, whether an action is taken or not does not affect the prediction, making this term zero, meaning no model exists. Thus, KL divergence, which represents model complexity, turns the binary distinction between model-free and model-based approaches into a continuous spectrum.



	<img alt="" class="aligncenter size-full wp-image-2399" height="317" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture4.png" width="605" />



	▷ a) Schematic of integrating habits and goal frameworks. b) Structure of the framework during training. c) Structure of the framework during behavioral processes.



	In this framework, when faced with a new goal, early learning resembles a model-based system, similar to a chef who, upon opening his restaurant, tries to optimize his predictions of new customers&#39; tastes. Once this predictive model has been sufficiently refined through continuous training, it may shift toward a more habit-driven approach, where the chef continually perfects his signature dishes.



	&nbsp;


<h2 style="line-height: 1.5;">
	3. Significance: Enabling AI to Perform Zero-Shot Learning
</h2>


	One significant advantage of human intelligence is the ability to solve various tasks in entirely new environments without relying on prior examples. For instance, when a painter is asked to depict a mythical creature like a qilin, he or she only needs to know that the qilin is a symbol of good fortune. However, an AI would require specific prompts, such as: &quot;Depict a qilin from ancient Chinese mythology, with a dragon&#39;s head, deer antlers, lion&#39;s eyes, tiger&#39;s back, bear&#39;s waist, snake&#39;s scales, horse&#39;s hooves, and an ox&#39;s tail, exuding an overall aura of majesty and sanctity, with gold and red as the primary colors, against a backdrop of swirling clouds, symbolizing good fortune, peace, and imperial power.&quot;



	The first painter to depict a qilin did so through zero-shot learning after acquiring sufficient experience in painting; however, zero-shot learning remains a challenging task for current AI systems. This is precisely the issue that this framework aims to address.



	When the environment changes, an agent built using the integrated framework proposed in this paper can spontaneously switch from habit-based model-free learning to model-based learning. This allows the agent to adapt to the new environmental conditions. In their experiments, researchers used a T-maze to test the adaptability of these agents. In this maze, the agent must decide which direction to take based on rewards on either side, learning strategies to maximize rewards.



	<img alt="" class="aligncenter size-full wp-image-2400" height="190" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture5.png" width="500" /> ▷ T-maze and the three stages of the agent in response to environmental changes.



	In a habit-based system, the agent might continue following the previously rewarded path, even if the reward has changed. However, goal-directed agents face a different challenge. For example, if the initial reward on the left side of the maze is 100 times greater than on the right, the agent might need to try the left side a hundred times before updating the model and attempting the right side (depending on the specific agent model). This is undoubtedly an inefficient approach; in the real world, an organism displaying such behavior would likely be eliminated by natural selection. The framework proposed in this paper integrates goal-driven and habit-driven approaches, allowing the agent to balance flexibility and speed. Initially, it adapts to the environment by choosing the right side; when the environment changes (the left reward disappears), it readjusts to choose the left side.



	The simple T-maze experiment shows that the new framework aligns with Yann LeCun&#39;s concept of the world model. LeCun emphasizes that a world model has a dual role of planning for the future and estimating missing observations, and it should be an energy-based model. In goal-directed behavior, this framework uses the current state, goal, and intended actions as inputs, and outputs an energy value to describe their &quot;consistency.&quot; It can be said that the agent&#39;s decision-making in the T-maze builds upon and relies on the world model envisioned by LeCun.



	There is a long road ahead from the simple T-maze to highly complex large language models. However, based on the theoretical framework described in this paper, we can observe some important similarities. For example, in training language models, we typically predict based on existing vocabulary, similar to scenarios without specific goals set during the training phase. The flexibility of goal-directed planning in this framework comes from its ability to break down any future goal into a series of sequential steps, predicting only the next observation. This method minimizes the discrepancy between goal-directed intentions and prior distributions, thereby compressing the search space and making the search process more efficient. This approach is also suitable for large models.



	Additionally, based on the KL divergence term in the framework, we can understand the hierarchical structure in predictive coding, where a hierarchical information processing method is used to reduce model complexity. Predictive coding theory also suggests that the brain learns to recognize patterns by filtering out information that can be predicted through natural world patterns, thereby reducing unnecessary data. This information processing strategy echoes the information bottleneck theory, showing how cognitive processes can be optimized by using lower-dimensional representations.



	Finally, this theoretical framework not only enhances our understanding of healthy brain function but also provides new perspectives for understanding and treating neurological disorders. For example, patients with Parkinson&#39;s disease often struggle with goal-directed planning abilities and rely more on habitual behaviors. This may be due to high uncertainty within goal-directed intentions. Research on how medical interventions or deep brain stimulation (altering internal states) and sensory stimulation (altering brain input) can reduce this uncertainty may provide methods to improve motor control in Parkinson&#39;s patients.



	Moreover, research on autism spectrum disorder (ASD) can also benefit from this theoretical framework. Individuals with autism often exhibit repetitive behaviors, which may be related to an overemphasis on model complexity in predictive coding, affecting cognitive behavioral flexibility when adapting to changing environments. Introducing some randomness to increase behavioral diversity could be a potential intervention strategy.



	&nbsp;



	&nbsp;

[:zh]

[:]

Who Says Intuition and Deliberation Are Incompatible? A New Perspective Based on Free Energy

[:en]Who Says Intuition and Deliberation Are Incompatible? A New Perspective Based on Free Energy[:]

[:en]In the field of neuroscience, numerous research papers on memory frequently mention the name Henry Molaison (H.M.). He was a patient who suffered from severe epilepsy. In 1953, he decided to undergo a surgical procedure aimed at treating his epilepsy, which involved the removal of brain tissue, including the hippocampus* (Figure 1). After the surgery, H.M. appeared quite normal in most respects: his personality, abstract thinking abilities, and reasoning skills were unaffected. He also retained memories from his childhood, clearly remembering the address of his childhood home. However, what puzzled people was that H.M. had almost entirely lost the ability to create new long-term memories. He could not remember the doctors who frequently saw him after the surgery, nor could he remember having just eaten.

*The hippocampus is a part of the brain's limbic system located beneath the cerebral cortex, playing roles in short-term memory, long-term memory, and spatial navigation.

H.M.'s experience sparked interest in the functional relationship between the hippocampus and memory. Nowadays, neuroscientists and psychologists have begun to recognize that there are multiple types of memory within our brain: episodic memories of past experiences, semantic memories of facts, and both short-term and long-term memories. But why do these memories exist in different forms? And where in the brain are these different types of memories stored? These questions remain mysteries to scientists.

<img class="aligncenter size-large wp-image-1756" src="https://admin.next-question.com/wp-content/uploads/2024/02/72cd94ef3d0ddfff932acdd2f85a77ca.png" alt="" width="396" height="212" />
▷ Figure 1: The hippocampus was removed from H.M. during surgery. Figure source: brainfacts.

Memory refers to the brain's ability to encode, store, and retrieve past experiences. The storage and generalization of past experiences are beneficial for animals in making adaptive decisions. For example, remembering the safe path to a specific water source and generalizing these memory cues can help identify environmental features that predict new water sources, such as the surrounding terrain and types of vegetation. Thus, this ability to recognize and understand current situations based on past experiences and predict future events, known as predictive memory, is particularly important for an organism's ability to adapt to its environment.

<img class="aligncenter size-full wp-image-1670" src="https://admin.next-question.com/wp-content/uploads/2024/02/1ccdb6a35dcc8347cf069b42ebc856cf.png" alt="" width="938" height="351" />
▷ Original paper: Organizing memories for generalization in complementary learning systems

A new theory, proven through experiments with artificial neural networks, suggests that the brain might classify memories based on their future utility. Specifically, predictive memories**(such as what to eat for breakfast each day or the route to take to work) are stored in the neocortex of the brain, aiding in better adaptation to changing environments and making effective decisions. Meanwhile, less useful memories, such as the taste of a special drink at a party, are stored directly in the hippocampus, without being consolidated in the neocortex.

Thus, the brain classifies memories based on their usefulness and predictability, optimizing memory reliability to help us cope with novel situations.

*Predictive memory implies that past experiences and memories provide a certain level of predictability about the occurrence or characteristics of future events. When the brain receives new information, it uses previously stored memories to recognize and understand the current situation and predict future events. This predictability helps individuals better adapt to their environment and make effective decisions.
<h2>1. Systematic Consolidation in Neural Network Models</h2>
Current classic research theories on memory storage across brain regions include the complementary learning systems hypothesis (Figure 2). This theory posits two complementary learning systems in the brain: the hippocampus and the neocortex. Located deep within the brain, the hippocampus is believed to be particularly active during the initial stages of learning, primarily responsible for the rapid encoding and storage of new information, crucial for the short-term memory of space, facts, and events. In contrast, the neocortex, part of the brain's outer layer, engages in the longer-term storage and integration of information. The learning process in the neocortex is slower, but it can handle more complex and abstract information, integrating it into long-term memory. These two systems work together in a dynamic learning and memory process, with the hippocampus rapidly encoding new information during the early stages of learning and then gradually transferring this information to the neocortex for more stable, long-term storage.

<img class="aligncenter size-large wp-image-1669" src="https://admin.next-question.com/wp-content/uploads/2024/02/3b721bb7bf2af2e83ab69b0673e4def9-1024x718.jpeg" alt="" width="1024" height="718" />
▷ Figure 2: The complementary learning process of the hippocampus and neocortex in memory storage. Figure source: researchgate.

Based on the complementary learning systems hypothesis, researchers have developed a neural network model called the "Teacher-Notebook-Student" for systematic consolidation (Figure 3), comprising three functionally different neural networks: the Teacher, Notebook, and Student networks. The Teacher network represents the environmental context of biological survival, providing experiential input; the Notebook network represents the hippocampus, capable of rapidly encoding all details of each experience provided by the environment; and the Student network represents the neocortex, predicting the information provided by the environment (Teacher) through experiences recorded in the hippocampus (Notebook).

In this neural network model, the "Teacher" is a linear feedforward network generating "input-output pairs" through fixed weights and additional output noise. The "Student" is a linear feedforward network of matching size with learnable weights. The "Notebook" is a sparse Hopfield network. During model training, the "Student" learns weights from a finite set of experiences containing signals and noise, aiming to minimize the squared difference between the "Teacher's" output and the "Student's" prediction. Furthermore, it is hypothesized that the main goal of the "Student" in this model is to optimize generalization performance, eventually predicting new "Teacher" outputs more accurately.

<img class="aligncenter size-large wp-image-1668" src="https://admin.next-question.com/wp-content/uploads/2024/02/8d91fe681cecf618ecb928e72660e9ac-1024x426.png" alt="" width="1024" height="426" />
▷ Figure 3: Teacher-Student-Notebook structure diagram. Figure source: Reference 1.

Researchers used this neural network model to simulate the dynamic process of memory and generalization, studying the actual impact and characteristics of systematic consolidation*. To optimize the Student's memory retrieval, researchers modeled the system as having infinite notebook playback. In each simulation, they generated experiential inputs from three different Teacher networks with varying predictabilities (high predictability, slight predictability, and unpredictability).

The results (Figure 4) show that although the generalization error of the Student network monotonically decreased for noise-free Teacher network inputs, noisy Teacher network inputs ultimately resulted in poor generalization performance of the Student network (red solid line). This indicates that in unpredictable environments, excessive systematic consolidation can severely reduce the model's generalization performance, leading to overfitting by the neocortex (Student) to unpredictable elements in the environment (Teacher).

This also implies that, in constructing neural network models, it is necessary to consider the level of systematic consolidation to ultimately enhance the model's generalization ability.

*Systematic consolidation is modeled as the plasticity within the synapses of the Student network, guided by the reactivation of the Notebook, akin to the hypothesized way hippocampal replay aids in systematic consolidation.

<img class="aligncenter size-large wp-image-1666" src="https://admin.next-question.com/wp-content/uploads/2024/02/1958888f6289735f14932c3ca5de5475-1024x307.png" alt="" width="1024" height="307" />
▷ Figure 4: Dynamics of Student generalization error, Student memory error, Notebook generalization error, and Notebook memory error during the optimization of Student memory. Figure source: Reference 1.
<h2>2. Generalization Optimization in Complementary Learning Systems (Go-CLS Model)</h2>
&nbsp;
<h3>Go-CLS Model</h3>
The Go-CLS Model, based on the Systematic Consolidation Neural Network Model, considers the adverse effects of systematic consolidation on model generalization ability in unpredictable environments observed in previous experiments. Researchers adjusted the level of systematic consolidation in the model. The fundamental hypothesis of this theory is that the degree of systematic consolidation is adaptively regulated. Once consolidation reaches a certain level, further consolidation should be halted to prevent weakening the model's generalization ability. According to the training results of the current Go-CLS model, under this setting, the Student (represented by the red solid line) can achieve nearly optimal generalization performance from the experiences of each Teacher.

<img class="aligncenter size-large wp-image-1666" src="https://admin.next-question.com/wp-content/uploads/2024/02/1958888f6289735f14932c3ca5de5475-1024x307.png" alt="" width="1024" height="307" />
▷ Figure 5: Dynamics of Student generalization error, Student memory error, Notebook generalization error, and Notebook memory error during the optimization of generalization performance. (b) High predictability (no noise); (d) Slight predictability (minor noise); (f) Unpredictability (high noise). Figure source: Reference 1.
<h3>Comparison with the Systematic Consolidation Neural Network Model</h3>
Upon comparing the aforementioned models, researchers found significant differences in their performance when processing Teacher data of varying predictabilities. In the standard systematic consolidation model, all memories are consolidated, but the generalization ability greatly varies with the predictability of the Teacher. For Teachers with lower predictability, this model exhibits a higher generalization error (Figure 6, left). In contrast, the Go-CLS model mitigates the detrimental effects of overfitting by halting before the Student can reach optimal memory performance. Moreover, as the predictability of the Teacher increases, both models show improvements in generalization and memory performance (Figure 6, right).

<img class="aligncenter size-large wp-image-1665" src="https://admin.next-question.com/wp-content/uploads/2024/02/b1cf189a9f25565de102f584313913dc-1024x315.png" alt="" width="1024" height="315" />
▷ Figure 6: Comparison of Student generalization performance changes with the Teacher's predictability level under the Systematic Consolidation Model and Go-CLS Model. Left: Systematic Consolidation Neural Network Model, Right: Go-CLS Model. Figure source: Reference 1.

It is evident that the Go-CLS model not only differs from the Systematic Consolidation Neural Network Model in aspects of retrograde amnesia* (i.e., the loss of past memories) but also in its approach to time-dependent generalization. After simulating hippocampal removal, the Systematic Consolidation Neural Network Model consistently shows a trend of gradually losing memories, implying that past memories are easier to forget. In contrast, the Go-CLS model can explain different types of memory loss phenomena, including those patterns that change over time. Regarding time-dependent generalization, when processing experiences of different predictabilities, both models produce a variety of generalization curves, with the maximum generalization performance increasing as the Teacher's predictability increases. However, only the Go-CLS maintains its time-varying generalization ability when the Student overfits. The Systematic Consolidation Neural Network Model might even lead to excessive generalization by the Student, resulting in highly inaccurate outputs.

*Retrograde Amnesia: The inability to recall events or information from before a specific point in time, while memories after this point are relatively preserved. This type of memory loss is typically caused by brain damage or other neurological disorders, affecting the individual's recall of past experiences and information. Retrograde amnesia may involve specific types of memory, such as semantic or episodic memory, rather than affecting all memory types. Researchers typically classify the dynamics of hippocampal forgetting based on whether the memory deficits are similar in recent and remote memories (flat retrograde amnesia), more pronounced in recent memories (graded retrograde amnesia), or absent in both recent and remote memories (no retrograde amnesia).
<h3>Supplement to the Theory of Complementary Learning Systems</h3>
The experimental results of the Go-CLS neural network model offer new insights into the complementary learning systems hypothesis. This hypothesis fundamentally posits the hippocampus and neocortex as fast and slow learning modules, respectively.

Researchers employed the Go-CLS neural network model to compare the generalization performance of the coupled Student-Notebook network (Figure 7, left) with that of the isolated Student (Figure 7, middle) and Notebook networks (Figure 7, right). In the Go-CLS model, the "Student" represents the neocortex, and the "Notebook" represents the hippocampus. The experimental groups of isolated Student and Notebook networks simulate scenarios where learning occurs using only the neocortex or the hippocampus, respectively.

<img class="aligncenter size-large wp-image-1664" src="https://admin.next-question.com/wp-content/uploads/2024/02/3a6433f05ee1b7a89e07c3b57da92d29-1024x406.png" alt="" width="1024" height="406" />
▷ Figure 7: Schematic diagram of learning system models: (a) Student-Notebook network; (b) Student network only; (c) Notebook network only. Figure source: Reference 1.

The isolated Student network model is limited in its generalization performance because it must learn online from each Teacher network's experiences without the ability to revisit or review past information (Figure 8, comparing orange and black curves). The results indicate that models including both networks perform better in generalization than those with only a Notebook. Furthermore, when the Teacher network provides moderately predictable data, the Student-Notebook coupled network model significantly outperforms the isolated Student network model in generalization. Therefore, under conditions of moderate data predictability and environmental noise, the Student-Notebook coupled network model exhibits superior generalization performance.

Moreover, the "replay function," or revisiting past learning experiences, can most effectively enhance the Student's learning capacity when the number of experiences available to the Student equals the number of adjustable parameters in the learning model. However, this approach risks overly adapting to old pieces of information, leading to inflexibility in responding to new situations, akin to the "double descent" phenomenon in machine learning. This phenomenon occurs when overfitting is most severe with a medium amount of data relative to the network size. To avoid this, the optimal Student-Notebook network model adjusts the amount of systematic consolidation based on the predictability of the Teacher, suggesting a similar regulatory mechanism in the brain for learning and memory based on experience predictability.

<img class="aligncenter size-full wp-image-1663" src="https://admin.next-question.com/wp-content/uploads/2024/02/12ab55cc976173269fb5e5f150c6de16.png" alt="" width="720" height="712" />
▷ Figure 8: Advantages of the complementary learning system. Figure source: Reference 1.
<h2>4. Summary</h2>
The Go-GLS model introduces a fascinating theoretical proposition: the brain might classify memories based on their predictability and future usefulness.
<img class="aligncenter size-large wp-image-1662" src="https://admin.next-question.com/wp-content/uploads/2024/02/b1bc04d8dfe81d1692d1bf178c15edef-1024x673.png" alt="" width="1024" height="673" />
▷ Figure 9: A story illustrating the conceptual difference between the Systematic Consolidation model and the Go-CLS model. Figure source: Reference 1.

Imagine a scenario where a young girl enjoys a day by the lake with her father. Among the day's experiences, predictable relations, like freshly picked strawberries tasting sweet, should be systematically consolidated to reinforce memories of predictable experiences. However, unpredictable co-occurrences, such as the color of the father's shirt coincidentally matching the strawberries, should not be consolidated in the neocortex but can still be stored in the hippocampus as part of the day's situational memory.

Additionally, the Go-GLS model has intriguing parallels with artificial intelligence research. It defines predictability through the optimal approximation error of teacher-student pairs, differing from real-world scenarios where achieving optimal student weights may not always be feasible. The model's concepts are applicable across a broader spectrum of models, including those with more complex student architectures and modern deep learning models. This necessitates exploring network architectures and learning rules that better balance memory and generalization.

However, the Go-GLS model currently focuses on simple supervised learning problems. Future efforts should address the optimal consolidation problem in more complex generalization scenarios, such as reinforcement learning and few-shot learning in large language models.[:]

How does the brain distinguish and store memories?

[:en]How does the brain distinguish and store memories? [:]

Meeting Reports

“Brain & Mind Lectures” on neuroaesthetics: How does the brain perceive the presence of beauty?

[:en]“Brain & Mind Lectures” on neuroaesthetics: How does the brain perceive the presence of beauty?[:zh]大脑如何感受美的存在？“Brain & Mind Lectures”共赏神经美学[:]

There have been a number of studies that have successfully decoded speech articulation and other motor signals from brain signals to restore a subject&#8217;s lost ability to speak. While effective, these decoders require neurosurgical access to the brain and are not applicable in most scenarios.
&nbsp;
A recent study published in Nature Neuroscience describes a novel non-invasive decoding approach that uses functional magnetic resonance imaging (fMRI) recordings to reconstruct continuous natural language from cortical representations of semantic meaning. This non-invasive brain-computer interface can be used to identify meaning in perceptions, imaginations, and silent videos, and to generate intelligible word sequences. The study demonstrates the feasibility of a non-invasive brain-computer interface for semantic reconstructions.
&nbsp;
On May 31st, young scientists from Tianqiao and Chrissy Chen Institute voluntarily organized the second session of “AI for Brain Science Journal Club”. Dr. Yuan Ze, a young scientist from the Peking University Sixth Hospital and also a member of the Chinese Academy of Sciences, expounded on the research. Please find below the essence from his interpretation as compiled by the NextQuestion editorial team.
&nbsp;
<h3>How to decode language without invasively decoding the brain?</h3>
&nbsp;
This study describes a novel decoder that uses non-invasive fMRI brain recordings to reconstruct arbitrary stimuli that a subject is hearing or imagining in continuous natural language.
&nbsp;
To compare word sequences to subjects&#8217; brain responses, the researchers trained an encoding model that predicted how subjects&#8217; brains would respond to phrases in natural language. The experiment recorded the fMRI BOLD responses of subjects&#8217; brains over a 16-hour period of listening to a narrative story and constructed an encoding model for each subject, which was then trained to predict the brain&#8217;s response based on the semantic features of the stimulus words.
&nbsp;
It turned out that the decoded word sequences not only captured the meaning of the stimuli, but even predicted the exact words and phrases.
&nbsp;
<h3>Where does the cerebral cortex register linguistic messages?</h3>
&nbsp;
The study divided the brain data into a classic language network, a combined parietal-temporal-occipital network, and a prefrontal network. After decoding from each network in each hemisphere individually, the researchers found that predictions of the decoders from each network in each hemisphere were significantly more similar to actual stimuli than random predictions.
&nbsp;
The researchers also calculated the time course of decoding performance for each network and found that most of the time points that were significantly decoded from the whole brain could be decoded from both the combined network and the prefrontal network. They similarly compared decoder predictions across networks and across hemispheres and found that the similarity between each pair of predictions was significantly higher than random ones. This suggests that cortical networks carry a great deal of redundant information, and that future brain-computer interfaces may be able to selectively record from the most accessible brain regions to obtain sound decoding performance.
&nbsp;
<h3>Applications: where can we apply the non-invasive language decoder?</h3>
·Imagined speech decoding: decoding based on brain activity during imagination 
·Cross-modal decoding: linguistic reconstruction of non-verbal tasks 
·Attention decoding: semantic representations are modulated by attention, and semantic decoders should thus selectively reconstruct the stimuli under attention 
·Privacy implications: an important ethical consideration of semantic decoding is that it may compromise mental privacy.
&nbsp;
<h3>Lessons learnt: where do all the data noises come from?</h3>
&nbsp;
The study also assessed whether decoding errors reflected random noise in brain recordings, model-setting errors, or both. The results found that model setting errors were the main source of decoding errors, in addition to random noise in training and testing data. Also, collecting more data may not significantly improve decoding performance.
<a href="https://www.nature.com/articles/s41593-023-01304-9">Read the paper on Nature Neuroscience</a>

[:en]There have been a number of studies that have successfully decoded speech articulation and other motor signals from brain signals to restore a subject's lost ability to speak. While effective, these decoders require neurosurgical access to the brain and are not applicable in most scenarios.

&nbsp;

A recent study published in Nature Neuroscience describes a novel non-invasive decoding approach that uses functional magnetic resonance imaging (fMRI) recordings to reconstruct continuous natural language from cortical representations of semantic meaning. This non-invasive brain-computer interface can be used to identify meaning in perceptions, imaginations, and silent videos, and to generate intelligible word sequences. The study demonstrates the feasibility of a non-invasive brain-computer interface for semantic reconstructions.

&nbsp;

On May 31st, young scientists from Tianqiao and Chrissy Chen Institute voluntarily organized the second session of “AI for Brain Science Journal Club”. Dr. Yuan Ze, a young scientist from the Peking University Sixth Hospital and also a member of the Chinese Academy of Sciences, expounded on the research. Please find below the essence from his interpretation as compiled by the NextQuestion editorial team.

&nbsp;
<h3>How to decode language without invasively decoding the brain?</h3>
&nbsp;

This study describes a novel decoder that uses non-invasive fMRI brain recordings to reconstruct arbitrary stimuli that a subject is hearing or imagining in continuous natural language.

&nbsp;

To compare word sequences to subjects' brain responses, the researchers trained an encoding model that predicted how subjects' brains would respond to phrases in natural language. The experiment recorded the fMRI BOLD responses of subjects' brains over a 16-hour period of listening to a narrative story and constructed an encoding model for each subject, which was then trained to predict the brain's response based on the semantic features of the stimulus words.

&nbsp;

It turned out that the decoded word sequences not only captured the meaning of the stimuli, but even predicted the exact words and phrases.

&nbsp;
<h3>Where does the cerebral cortex register linguistic messages?</h3>
&nbsp;

The study divided the brain data into a classic language network, a combined parietal-temporal-occipital network, and a prefrontal network. After decoding from each network in each hemisphere individually, the researchers found that predictions of the decoders from each network in each hemisphere were significantly more similar to actual stimuli than random predictions.

&nbsp;

The researchers also calculated the time course of decoding performance for each network and found that most of the time points that were significantly decoded from the whole brain could be decoded from both the combined network and the prefrontal network. They similarly compared decoder predictions across networks and across hemispheres and found that the similarity between each pair of predictions was significantly higher than random ones. This suggests that cortical networks carry a great deal of redundant information, and that future brain-computer interfaces may be able to selectively record from the most accessible brain regions to obtain sound decoding performance.

&nbsp;
<h3>Applications: where can we apply the non-invasive language decoder?</h3>
·Imagined speech decoding: decoding based on brain activity during imagination
·Cross-modal decoding: linguistic reconstruction of non-verbal tasks
·Attention decoding: semantic representations are modulated by attention, and semantic decoders should thus selectively reconstruct the stimuli under attention
·Privacy implications: an important ethical consideration of semantic decoding is that it may compromise mental privacy.

&nbsp;
<h3>Lessons learnt: where do all the data noises come from?</h3>
&nbsp;

The study also assessed whether decoding errors reflected random noise in brain recordings, model-setting errors, or both. The results found that model setting errors were the main source of decoding errors, in addition to random noise in training and testing data. Also, collecting more data may not significantly improve decoding performance.

<a href="https://www.nature.com/articles/s41593-023-01304-9">Read the paper on Nature Neuroscience</a>[:zh]谈到脑机接口，通过侵入方式记录大脑信号，解读人类的所思所想已不是什么新鲜事。

已有不少研究成功实现了从大脑信号中解码语音发音和其他运动信号，来恢复受试者已经丧失了的说话能力。虽然有效，但这些解码器都要通过神经外科手术接入大脑，并不适用于大多数场景。那么，非侵入的方式有用吗？

以往，使用非侵入性记录的解码器（non-invasive decoder）只能从一小组字母单词或短语中识别刺激，一直具有较大的应用局限性。而近日，一篇发表在Nature Neuroscience上的研究介绍了一种新型的非侵入式解码方式，它使用功能性磁共振成像(fMRI)记录从语句意义的皮层表征中重建连续的自然语言。这种非侵入式的脑机接口可用于识别感知、想象和沉默视频中的意义，并生成可理解的单词序列。研究证明了非侵入式语言大脑 - 计算机接口的可行性。

&nbsp;

[caption id="attachment_805" align="aligncenter" width="600"]<img class="wp-image-805" src="https://admin.next-question.com/wp-content/uploads/2023/11/01.png" alt="" width="600" height="304" /> ▷图注：论文封面[/caption]

&nbsp;

2023年5月31日，天桥脑科学研究院（Tianqiao and Chrissy Chen Institute，TCCI）的青年科学家们自发组织了 “AI for Brain Science Journal Club”第二期会议，来自中国科学院大学北京大学第六医院的青年科学家袁则博士详细解读了这项研究，追问媒体对会议内容进行梳理，以飨读者。

&nbsp;
<h3>不侵入大脑，如何解码语言？</h3>
&nbsp;

这项研究介绍了一种新型解码器，它采用非侵入性fMRI大脑记录，并能以连续的自然语言重建受试者正在听到或想象的任意刺激。为了将单词序列与受试者的大脑反应进行比较，研究者训练出一个编码模型，预测受试者的大脑如何对自然语言中的短语做出反应。试验记录了3名受试者在听叙事故事16小时内大脑的fMRI BOLD反应，并以此为每个受试者构建编码模型，然后训练该模型，使其能够根据刺激词的语义特征预测大脑的反应。编码器将大脑反应输入到解码器，再由解码器将其翻译为一组候选单词序列并进行评分，并保留最有可能的单词序列（波束搜索算法）（图1）。该语义解码器使用的波束搜索算法能有效改善fMRI低时间分辨率对预测结果准确度的影响。

[caption id="attachment_808" align="aligncenter" width="600"]<img class="wp-image-808" src="https://admin.next-question.com/wp-content/uploads/2023/11/04.png" alt="" width="600" height="198" /> ▷图1：语义解码器的训练流程图 图片来源：Nature Neuroscience[/caption]

&nbsp;

结果发现，解码出的单词序列不仅捕获了刺激的含义，甚至预测了精确的单词和短语。

[caption id="attachment_809" align="aligncenter" width="600"]<img class="wp-image-809" src="https://admin.next-question.com/wp-content/uploads/2023/11/05.png" alt="" width="600" height="200" /> ▷图2：志愿者听到的语句（左）和解码器根据大脑活动解读的语句（右），蓝色代表完全一致的词汇，紫色代表大意准确的词汇[/caption]

&nbsp;

为了量化解码性能，研究还使用几个“语言相似性”度量来比较一则测试故事（1800个单词）的解码和实际单词序列（图3）。通过一系列语言相似性的度量，解码器所预测结果与实际单词的相似性、解码分数、识别准确率显著高于偶然与随机。

[caption id="attachment_810" align="aligncenter" width="600"]<img class="wp-image-810" src="https://admin.next-question.com/wp-content/uploads/2023/11/06.png" alt="" width="600" height="202" /> ▷图3：语义解码器解码性能量化。[/caption]

&nbsp;
<h3>语言信息藏在大脑皮层何处？</h3>
&nbsp;

为了回答哪些皮层网络代表了足够详细的语言，以及不同的网络（或半球）在语言处理中是互补还是冗余的等问题，研究将大脑数据划分为三个皮层网络：经典语言网络、顶叶-颞-枕叶联合网络和前额叶网络（图4）。研究者从每个半球的每个网络中单独解码后发现，来自每个半球的每个网络的解码器预测与实际刺激的相似性显著高于随机预期。

[caption id="attachment_811" align="aligncenter" width="600"]<img class="wp-image-811" src="https://admin.next-question.com/wp-content/uploads/2023/11/07.png" alt="" width="600" height="203" /> ▷图4：皮层网络划分[/caption]

&nbsp;

研究者还计算了每个网络解码性能的时间过程，发现从整个大脑中显著解码的大多数时间点都可以从联合网络（77%-83%）和前额叶网络（54%-82%）中进行解码（图5）。他们同样比较了跨网络和跨半球的解码器预测，发现每对预测之间的相似性显著高于随机。这表明，这些皮层网络承载了大量冗余信息，未来脑-机接口或许可以选择性地从最容易接近的大脑区域进行记录（而非特定的脑区）来获得良好的解码性能。

[caption id="attachment_812" align="aligncenter" width="600"]<img class="wp-image-812" src="https://admin.next-question.com/wp-content/uploads/2023/11/08.png" alt="" width="600" height="198" /> ▷图5：语义解码器用于不同皮层网络的解码效果[/caption]

&nbsp;
<h3>应用：非侵入性语言解码器用在哪里？</h3>
&nbsp;

为了探索这种新型解码器的应用价值，研究者在故事预测期间使用大脑反应为每个受试者训练了单个语义语言解码器，然后将其应用于其它任务期间的大脑反应上。

&nbsp;

· 想象语音解码：根据大脑想象过程中的活动进行解码。针对每一个故事，将解码器根据受试者想象预测的故事与受试者在不进行扫描时描述的故事进行比较，正确地识别出了哪个预测结果对应于哪个故事（100%正确率）。

· 跨模态解码：针对非语言任务进行语言重建。使用fMRI记录受试者观看了四部没有声音的短片时的大脑活动并使用语义语言解码器进行解码。将解码后的单词序列与针对视障人士的电影音频描述进行比较，发现解码的序列准确地描述了来自电影的事件。这表明，在故事感知过程中训练的单个语义解码器可以用于解码一系列语义任务。

· 注意力解码：语义表征受到注意力的调节，照理语义解码器应该有选择地重建被关注的刺激。为了测试这一点，实验对象听了两次重复的多个说话者刺激，该刺激是通过暂时叠加由女性和男性说话者讲述的两个故事来构建的。在每次演讲中，受试者都被提示去听不同的演讲者。解码器的预测与受试者描述的故事表现一致，表明解码器选择性地重构了被关注的刺激。

· 隐私影响：语义解码的一个重要的伦理考虑是它可能会损害精神隐私。研究试图使用根据其他受试者的数据训练的解码器来解码每个受试者的感知语音。结果表明，受试者合作对于解码器训练仍然是必要。而且，语义解码可以被有意识地抵制。

&nbsp;
<h3>借鉴：数据噪声从何而来？</h3>
&nbsp;

为了进一步改进解码器的解码效果，研究还评估了解码错误是否反映了大脑记录中的随机噪声、模型设定错误还是两者兼而有之。结果发现，除了训练和测试数据中的随机噪声之外，模型设定错误是解码错误的主要来源。

为了评估解码性能是否受到训练数据集大小的限制，研究使用不同数量的数据来训练解码器。虽然解码性能随着训练数据量的增加而提高，大多数改进发生在第七次扫描会话时，即7.5小时，这表明简单地收集更多数据可能不会显著提高解码性能。

此外，测试数据中的低信噪比（SNR）也可能会限制可以解码的信息量。研究发现，通过对不同重复测试故事期间收集的大脑反应进行平均来人为增加信噪比，解码性能随着平均响应的数量而略微增加，这表明解码误差的一些分量反映了测试数据中的噪声。值得注意的是，解码性能与训练刺激中的词频没有显著相关性，这表明模型的错误设定不是主要由训练数据中的噪声引起的。此外，研究者还发现解码性能与单词具体性的行为评级显著相关，这表明解码器在重建具有某些语义属性的单话方面较差。

&nbsp;
<h3>会议追问</h3>
&nbsp;

会议最后，袁则博士总结道，这项研究表明感知和想象刺激的意义可以从fMRI记录中解码为连续的语言，这标志着非侵入性脑机接口的重要一步。虽然解码器成功地重建了语言刺激的意义，但它经常无法重建准确的单词且可能会发生特异性的损失。另外，主体反馈是提高解码性能的重要因素，这种反馈允许受试者适应解码器，为他们提供对解码器输出的更多控制。

&nbsp;
<h4>Q：研究中主要使用了两种方法进行解码，可以具体介绍一下这两种方法吗？</h4>
A：一是逐词生成候选序列的“波束搜索算法”。在波束搜索中，当基于听觉和语言区域的大脑活动检测到新单词时，语言模型为波束中的每个候选序列生成延续。然后，编码模型对每次延续诱发记录的大脑反应的可能性进行评分，最有可能的延续被保留在下一时间步的波束中；另一种方法是GPT，预训练生成的GPT是一个12层的神经网络，它使用多头自我注意机制将序列中每个单词的表示与之前单词的表示相结合。GPT在一个大的书籍语料库上被训练来预测下一个单词的概率分布。

&nbsp;
<h3>会议海报</h3>
<img class="aligncenter wp-image-813" src="https://admin.next-question.com/wp-content/uploads/2023/11/09.png" alt="" width="600" height="1067" />

&nbsp;

<a href="https://www.nature.com/articles/s41593-023-01304-9">阅读原文</a>[:]

How do you read the mind without invasively decoding the brain?

[:en]How do you read the mind without invasively decoding the brain?[:zh]不侵入大脑，如何解读心声？[:]

Is AI Dream Interpretation Reliable? Current Progress in Academia

[:en]Is AI Dream Interpretation Reliable? Current Progress in Academia[:]

When we discuss &#8220;intelligent animals,&#8221; which species come to mind? Perhaps it&#8217;s the Border Collie, epitomized by the saying, &#8220;Border Collies are to dogs what geniuses are to humans,&#8221; the gorilla, our &#8220;cousin&#8221; in the animal kingdom, or the orca, celebrated for its collaborative hunting strategies and instances of rescuing attacked divers.
In Australia, however, the sulfur-crested cockatoo might top the list.
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1621" src="https://admin.next-question.com/wp-content/uploads/2024/02/e174de97e868a62e0925b9ad67f6ea54.gif" alt="" width="600" height="338" /> 
▷ Sulfur-crested cockatoo skillfully opens the bin. Source: livescience.
These birds, known for their striking white plumage and sunflower-like yellow crests, are a common sight in zoos around the world, thanks to their successful breeding in captivity and remarkable learning abilities. Yet, in their native Australian environment, they have become a nuisance to residents. The cockatoos have mastered opening various types of garbage bins to forage for food, spreading this skill within their communities through social learning.
A study published in Science highlighted how the sulfur-crested cockatoos&#8217; bin-opening behavior rapidly proliferated, spreading from three to 44 Sydney suburbs from before 2018 to 2019. The researchers noted regional differences in the opening techniques, which varied by area, often separated by forests [1]. These ingenious birds have compelled humans to invent new technologies to secure their garbage bins, showcasing a fascinating interplay between humans and wildlife.
Interestingly, when considering &#8220;intelligence,&#8221; many overlook birds, favoring mammals, particularly those closely related to us. Yet, reflecting on examples from &#8220;parrots mimicking speech&#8221; to &#8220;crows utilizing tools for water displacement,&#8221; it&#8217;s evident that birds&#8217; intelligence is deeply ingrained in our cultural consciousness. Behavioral studies confirm that corvids and parrots exhibit cognitive skills across various domains comparable to those of non-human primates [2][3]. These include spatial and episodic memory, understanding cause and effect, delaying gratification, planning for the future, and even possessing a theory of mind.
What accounts for the remarkable intelligence of these avian species?
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1620" src="https://admin.next-question.com/wp-content/uploads/2024/02/afab9d271755a8cb0739c2a83e880274.png" alt="" width="1016" height="292" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/afab9d271755a8cb0739c2a83e880274.png 1016w, https://admin.next-question.com/wp-content/uploads/2024/02/afab9d271755a8cb0739c2a83e880274-300x86.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/afab9d271755a8cb0739c2a83e880274-768x221.png 768w" sizes="(max-width: 1016px) 100vw, 1016px" /> 
▷ Original paper: Güntürkün, Onur, Roland Pusch, and Jonas Rose. &#8220;Why birds are smart.&#8221; Trends in Cognitive Sciences (2023).https://doi.org/10.1016/j.tics.2023.11.002 
&nbsp;
<h2>1. &#8220;Size&#8221; Doesn&#8217;t Matter, &#8220;Content&#8221; Is Key</h2>
To grasp the concept of intelligence, one must delve into the study of the brain. The intelligence exhibited by birds overturns a longstanding assumption that a larger brain equates to greater intelligence. While scientists cannot solely judge the intelligence of a species or an individual by the weight of its brain, within a given taxonomic group, a brain that is larger and heavier typically contains more neurons and offers greater computational capabilities, rendering its possessor more intelligent compared to its &#8220;cousins.&#8221; [4] This principle holds true for non-human primates, cetaceans, and birds. However, when comparing species that are more distantly related, the absolute and relative sizes of the brain are less indicative of intelligence. For example, corvids and parrots, with their brains weighing merely 5-20 grams, can rival the intellectual abilities of gorillas, whose brains weigh around 400 grams.
What enables these birds with walnut-sized brains to achieve such feats of intelligence?
A study published in PNAS in 2016 shed light on this question, revealing that despite their small size, birds&#8217; brains boast an exceptionally high neuron density [5]. In fact, birds have twice as many neurons per unit volume as primates and four times as many as mice. This dense neuron population allows the avian forebrain to possess a number of neurons comparable to that found in primates, laying the groundwork for similar information processing capabilities. Neurons, the fundamental units of brain computation, are crucial for determining a brain&#8217;s computational capacity. Birds, particularly with their higher proportion of pallial neurons, exhibit advanced cognitive functions. For instance, the corvus frugilegus and the callithrix jacchus have comparable brain weights, yet the corvus frugilegus possesses three times as many pallial neurons, which are essential for integrating various cognitive processes towards achieving complex goals. This abundance of pallial neurons makes the corvus frugilegus more intellectually capable than the callithrix jacchus.
Another neuron type crucial for cognitive flexibility is associative neurons, which mediate between sensory inputs and motor outputs, facilitating associative and motor learning. The New Caledonian crow has a significant number of associative neurons, comparable to those found in the prefrontal cortex of gorillas, further highlighting its sophisticated intelligence [6]. This crow species adeptly uses sticks to extract beetle larvae from tight spaces, showcasing its superior problem-solving skills. 
In summary, the remarkable computational abilities of the small bird brain, underscored by high neuron density, a significant number of pallial neurons, and associative neurons, provide a physiological foundation for avian intelligence. 
&nbsp;
<h2>2. The &#8220;Sufficient but Not Necessary&#8221; Neocortex</h2>
Birds and mammals embarked on separate evolutionary paths hundreds of millions of years ago, leading to significant differences in their forebrain structures. In mammals, the dorsal pallium evolves into the cerebral cortex, predominantly forming the isocortex. The term &#8220;isocortical&#8221; describes areas of the cerebral cortex that are structurally similar. This isocortex, also known as the &#8220;neocortex,&#8221; is unique to mammals, representing an evolutionary advancement in complexity and functionality. The neocortex encompasses regions responsible for perception, motor functions, and cognitive associations, playing a pivotal role in cognition and learning abilities. Traditionally, the neocortex was viewed as the cornerstone of mammalian intelligence. Yet, the intelligence displayed by birds prompts a reevaluation of this perspective.
In birds, the dorsal pallium—a structure analogous to that in mammals—develops into the hyperpallium. The avian hyperpallium, primarily involved in perceptual functions, differs significantly from the mammalian neocortex and does not directly correspond in terms of functionality. Beneath the lateral ventricles lie most of the avian pallial nuclei, collectively known as the dorsal ventricular ridge (DVR). The DVR, lacking a mammalian equivalent, functionally complements the hyperpallium by processing perceptual information and incorporating motor and associative regions. Recent research indicates that certain DVR regions involved in processing perceptual information exhibit a layered structure and information processing pathways akin to those found in the mammalian cerebral cortex [7]. However, the motor and associative areas maintain a nuclear arrangement.
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1619" src="https://admin.next-question.com/wp-content/uploads/2024/02/71f8a1c998450ccf362887b06cb14f0f-1024x774.png" alt="" width="1024" height="774" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/71f8a1c998450ccf362887b06cb14f0f-1024x774.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/02/71f8a1c998450ccf362887b06cb14f0f-300x227.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/71f8a1c998450ccf362887b06cb14f0f-768x581.png 768w, https://admin.next-question.com/wp-content/uploads/2024/02/71f8a1c998450ccf362887b06cb14f0f.png 1103w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ A) The segmentation of the bird brain; B) The names of structures in the bird brain (highlighted in yellow) and their homologous structures in the mammalian brain (in white); C) The different information processing pathways, with C1 representing the isocortex, C2 the dorsal ventricular ridge, and C3 the hyperpallium. Source: Original paper.
These findings indicate that convergent evolution has led to similarities between the avian and mammalian brains in structure and information processing techniques. In terms of perceptual information processing, these isocortex-like structures may provide indispensable advantages. Nonetheless, given the evolutionary divergence between the DVR and neocortex and the distinct processing methods of the DVR, the neocortex appears to be a &#8220;sufficient but not necessary&#8221; condition for intelligence. 
&nbsp;
<h2>3. The Dopamine-Driven &#8220;Prefrontal Cortex&#8221; of Birds</h2>
We have previously examined the intelligence of birds, focusing on neuron numbers, types, and brain structures, and have noted both similarities and differences with mammals. We will now delve further into the reasons behind avian intelligence by exploring brain function.
The nidopallium caudolaterale, located at the end of the dorsal ventricular ridge, plays a crucial role akin to the mammalian prefrontal cortex, encompassing a wide range of cognitive processes. Notably, this region and the mammalian prefrontal cortex share the highest density of dopaminergic neurons in the brain, linking them to associative areas and premotor structures. Neurons in the nidopallium caudolaterale, like those in the mammalian prefrontal cortex, are adept at categorizing perceptual information (e.g., size and quantity), organizing stimuli in varying sequences (either prospectively or retrospectively), and encoding executive functions and sensory awareness. This similarity underscores the nidopallium caudolaterale&#8217;s pivotal role in orchestrating complex cognitive tasks, thereby exhibiting intelligent behavior in birds. 
&nbsp;
<h2>4. Do Birds Have Working Memory?</h2>
Working memory, the capacity to hold information for real-time processing, is essential for cognitive functions and is primarily associated with the prefrontal cortex. Research involving brain injury, pharmacological studies, and neurophysiology indicates that the nidopallium caudolaterale in birds also supports working memory. Techniques like single-cell recordings have identified &#8220;delay activity&#8221; in this area, mirroring the mammalian prefrontal cortex. This activity, where neurons remain active during the interval between stimulus presentation and response, signifies the presence of working memory in birds [8]. Furthermore, neurophysiological patterns akin to those in mammals, derived from the collective activity of multiple neurons, have been observed in birds, suggesting that despite their structural differences, avian cognitive activities can mimic mammalian neurophysiological signatures.
Studies on working memory often concentrate on individual neurons and specific brain regions. However, research on sleep and dreaming offers insights into cognitive abilities at a whole-brain level. For instance, experiments on pigeons during rapid eye movement (REM) sleep—a phase closely associated with dreaming—have shown brain activation patterns involving the limbic system, premotor regions, and visual and multimodal areas, similar to humans [9]. Although we cannot yet decipher dream content from such activity, there is speculation that pigeons may dream about navigating obstacles in flight. These findings suggest that birds&#8217; cognitive processes involve extensive neural network activities comparable to those in mammals.
Despite these similarities, the neurophysiological underpinnings of avian cognition also display unique characteristics. For example, unlike mammals, birds do not seem to use the hippocampus for memory consolidation during sleep, raising questions about how they manage and access long-term memories—a topic ripe for further investigation. 
&nbsp;
<h2>5. Why Are Birds So Intelligent?</h2>
In their latest opinion article in Trends in Cognitive Sciences, the authors explore the question, &#8220;Why are birds so intelligent?&#8221; to understand what constitutes a &#8220;smart brain.&#8221; So, what characteristics enable a brain to develop intelligence?
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1618" src="https://admin.next-question.com/wp-content/uploads/2024/02/db68bf4dda445d9ae025f2af71d9855a-1024x710.png" alt="" width="1024" height="710" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/db68bf4dda445d9ae025f2af71d9855a-1024x710.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/02/db68bf4dda445d9ae025f2af71d9855a-300x208.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/db68bf4dda445d9ae025f2af71d9855a-768x533.png 768w, https://admin.next-question.com/wp-content/uploads/2024/02/db68bf4dda445d9ae025f2af71d9855a.png 1280w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ Figure: Venico_J
Firstly, intelligence is not merely a function of brain size. The quantity, types, and interconnectedness of neurons within the brain are, arguably, more vital. Birds, despite having smaller brains compared to mammals, boast a high density and large number of cortical and associative neurons, enabling them to achieve intelligence within a compact brain size.
Secondly, possessing a mammalian-specific neocortex is not essential for intelligence. The avian dorsal ventricular ridge (DVR), though not evolutionarily homologous to the mammalian neocortex and differing in physiological structure and processing pathways, performs similar functions in processing perceptual, motor, and associative information. This indicates that the neocortex is &#8220;sufficient but not necessary&#8221; for intelligence development.
Thirdly, the nidopallium caudolaterale in birds, rich in dopaminergic neuron connections, plays a role in nearly all cognitive processes. Functionally akin to the mammalian prefrontal cortex, it integrates information, encodes abstract cognition, and coordinates actions, suggesting that a control center resembling the prefrontal cortex is pivotal for complex cognitive abilities. This resemblance may be a result of convergent evolution.
Lastly, research on avian working memory and dreams during sleep demonstrates that, despite structural differences, avian cognitive processes share neurophysiological similarities with those of mammals. This suggests that the mechanisms underlying neuron and neural network cooperation in intelligence development share fundamental, indispensable commonalities.
By examining birds, we challenge the exclusive association of intelligence with mammals, exploring the concept of a &#8220;smart brain&#8221; from the perspectives of neuron numbers and types, brain structure, function, and neurophysiological activity. These are merely the directions currently well-explored, with other potential areas awaiting discovery.
The study of avian brains not only disrupts the notion that intelligence is unique to mammals but also offers a fresh lens through which we can examine our own intelligence and the essence of life itself. In questioning &#8220;Why are birds so intelligent?&#8221; we delve into our understanding of intelligence and the boundaries of life.

Why Are Birds So Intelligent?

[:en]Why Are Birds So Intelligent?[:zh]鸟儿为什么这么聪明？[:]

The Brain: The Universe&#8217;s Most Complex System. For a long time, humans have endeavored to unravel the brain&#8217;s inner mechanisms, yet progress has been somewhat sluggish, constrained by technological and methodological limitations. However, with the advent of artificial intelligence technology, unprecedented opportunities have emerged for brain science research.
On June 21st, the Tianqiao and Chrissy Chen Institute (TCCI), in collaboration with Neurochat, initiated the fourth TCCI-Neurochat Online Academic Conference. On the opening day, Professor Xiao-Jing Wang from New York University delivered a keynote address titled &#8216;Theoretical Neuroscience in the Age of Artificial Intelligence&#8217;. He shared insights into the integration of artificial intelligence with neuroscience from a theoretical perspective, aiming to advance research in the field.
Theoretical or Computational Neuroscience involves applying mathematical theories to understand how the brain operates. Similar to the role of theoretical physics in the field of physics, theoretical neuroscience plays a crucial role in brain research, complementing experimental studies. Experimentation alone may not fully reveal how the brain functions on multiple levels. Theoretical neuroscience transforms neurons and neural networks into mathematical models and algorithms, thus building a bridge between brain science and AI. With advancements in experimental tools and the advent of big data, theoretical research and modeling have become increasingly essential.
&nbsp;
<h2>Brain Functions Beyond AI’s Reach</h2>
The technological revolution in the AI field is often traced back to the year 2012. That year, Geoffrey Hinton’s research team demonstrated that deep neural networks could perform visual object recognition tasks. It&#8217;s well-known that different cortical areas of the brain correspond to different functions. Similarly, the various layers of deep neural networks are often analogized to brain regions such as V1, V2, and V4. However, the prefrontal cortex (PFC), situated just above the eyes, is still not sufficiently understood. Additionally, a functional equivalent of the PFC in contemporary machine systems has yet to be determined.
This enigmatic PFC region is crucial for various cognitive functions, including working memory and decision-making. Working memory refers to the brain&#8217;s capacity to maintain and manipulate information internally, thus enabling focus on inner thoughts without external distractions.  Decision-making involves making choices under uncertainty.
In laboratories, working memory is typically studied using delay-dependent tasks. For example, a monkey is shown the location of food and then subjected to a delay phase of 5 or 10 seconds. During this delay, the monkey must remember whether the food was placed on the left or the right. After the delay, the monkey retrieves the food based on memory, not direct sensory stimuli (the food itself). In the early 1970s, using single-cell recording and electrical stimulation, it was discovered that individual PFC cells remained active during the delay phase. These neurons were not active before the stimulus and did not react to the stimulus itself but showed continued activity during the delay, which intensified as the delay period doubled. This phenomenon is known as &#8216;stimulus-selective persistent activity&#8217;.
<div id="attachment_1405" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1405" class="wp-image-1405" src="https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1-1024x768.png" alt="" width="600" height="450" srcset="https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1-1024x768.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1-300x225.png 300w, https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1-768x576.png 768w, https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Delay-Dependent Task. Figure courtesy of Professor Xiao-Jing Wang.</div>
In the study of decision-making, a simple yet ingenious experiment stands out: the two-alternative forced choice. Monkeys are trained to fixate on a point on a screen, and upon its disappearance, they make a decision: determining whether the movement of the point was to the left or to the right. Their decision is expressed through eye movements. The trick lies in the fact that, during any given trial, the proportion of the point&#8217;s movement direction, known as coherence or motion strength, is controllable. If all points move in one direction, the decision is easy. But if only 50% move in the same direction and the rest move randomly, it becomes more challenging. And with a mere 5% coherence, or even 0%, making a subjective judgment becomes exceedingly difficult. Single-cell physiological studies have shown that neurons exhibit ramping activity within milliseconds. This instantaneous and intense activity represents a mechanism through which neurons integrate information, accumulating evidence about different choice options to form a subjective judgment.
Professor Xiao-Jing Wang, through these two experimental examples, strikes at the core of the most complex issues. As the renowned psychologist Donald Hebb once said, &#8216;If one were to explain the delay between stimulus and response with a central neural mechanism, it seems remarkably congruent with the characteristics of thinking.&#8217; When we think, it&#8217;s not driven by direct external stimuli but is an internal brain activity. Delayed activity is a specific instance of internal neuronal group activity.
However, the biological basis of decision-making or choice remains unclear, and AI systems lack this capability. As Noam Chomsky notes, &#8216;Once you get to issues of will or decision or reason or choice of action, human science is at a loss.&#8217;
&nbsp;
<h2>Modeling Based on Biology</h2>
Despite the complexity of real biological processes, relatively simple behavioral paradigms can be used to attempt understanding mechanisms across different levels.
In a physics experiment involving a pendulum, if there is friction in the air, the pendulum&#8217;s amplitude decreases over time, exhibiting an exponential decay. Similarly, neuronal firing also decays over time, with a time constant of approximately 10 milliseconds. To achieve sustained activity in working memory, this issue needs to be addressed. Professor Wang proposes a bold hypothesis: reverberation. That is, if a group of neurons stimulates each other, they can continue to be active even after the input disappears. The strong horizontal connections in layers two and three of the PFC in primates provide a structural basis for this hypothesis.
To test this hypothesis, Professor Wang simplified the complexity by creating a system with only two dynamic variables. This system demonstrates how to integrate random dots, perceptual decision-making, and working memory for judgments and choices.
This model has two selective firing neuronal groups &#8211; a and b. There is reverberation within each neuronal group, but the excitatory selectivity between the two groups is much weaker. They compete through inhibitory neurons, ultimately resulting in a winner. This outcome represents the system&#8217;s choice. The system&#8217;s choice is determined by the motion strength and relative input to groups a and b. When the input disappears, the choice signal remains self-sustained throughout the delay period. This model can explain the basic processes of working memory and decision-making.
<div id="attachment_1404" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1404" class="wp-image-1404" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png" alt="" width="600" height="452" srcset="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-300x226.png 300w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-768x578.png 768w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />Caption: (Left) Two-Alternative Forced Choice; (Right) Slow Reverberation Dynamics System. Figure courtesy of Professor Xiao-Jing Wang.</div>
Furthermore, this model operates across multiple levels. By considering reaction time as a function of motion strength, it is observed that the decision-making time increases as the motion strength decreases and the task becomes more difficult. This mirrors our everyday experience, where making a tough choice often requires considerable time to weigh different options. This phenomenon is also reflected in monkey experiments, elucidating why the dynamics of recurrent neural circuits can explain behavior.
Additionally, Professor Wang offers insights from a cellular and molecular perspective. For instance, slow time integration is primarily achieved through excitatory reverberation mediated by NMDA receptors. NMDA receptors are crucial for decision-making and working memory, as tested and confirmed in monkey experiments. If a drug were found that reduces or eliminates NMDA signaling, then neuronal activity would cease.
When considering self-sustained and internal states as functions of recurrent activation strength, and labeling these states with a parameter, it is found that in this model, above a certain parameter threshold, a series of stimulus-selective, persistent states emerge. These states include different memory items in working memory. The emergence of these states can be explained as bifurcations in the dynamic system, where graded changes in parameters lead to different functions and capabilities. This perspective is vital for recurrent neural circuits and is also applicable in multi-regional, large-scale brain systems.
<div id="attachment_1403" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1403" class="wp-image-1403" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1024x794.png" alt="" width="600" height="465" srcset="https://admin.next-question.com/wp-content/uploads/2024/01/640-1024x794.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/01/640-300x233.png 300w, https://admin.next-question.com/wp-content/uploads/2024/01/640-768x595.png 768w, https://admin.next-question.com/wp-content/uploads/2024/01/640.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />Caption: Sufficiently Strong Recurrent Connections in Working Memory and Decision-Making. Figure courtesy of Professor Xiao-Jing Wang.</div>
Building upon this foundation, we can utilize machine learning to establish neural circuits akin to biological ones, capable of performing 20 different tasks, including working memory, decision-making, categorization, multisensory integration, delay sampling, delay categorization, and more. Once the network successfully executes these nuanced tasks, it enables us to &#8216;open the black box&#8217; for analysis, pose questions, and delve into the dynamics of new circuits and neuronal groups. Neuronal activity can be gauged through task differentiation, and the interactions and organizational patterns among neurons can be deciphered using cluster analysis.
However, even if we design circuits that resemble the PFC, capable of flexibly guiding behavior based on rules, our endeavor should not end there. In reality, any given function involves multiple brain regions, including the PFC, the posterior parietal cortex, and potentially even supra-cortical structures.
&nbsp;
<h2>From Local Networks to Brain Systems</h2>
The interactions between different regions of the brain exhibit a high degree of recursion. As a theoretician, Professor Wang asserts that merely qualitatively sketching their feedback connections is insufficient; he advocates for the clarity and precision of numerical descriptions. Over the past decade, technological advancements have made it feasible to utilize collected data for constructing directed and weighted connection matrices between various regions.
In the cortex of macaque monkeys, when the directed connection weights between two regions are represented in histograms, it is observed that these histograms across multiple regions follow a log-normal distribution. The weight of any pair of regions and their wiring distance exhibit an exponential distribution. This implies that traditional graphical representations of cortical networks fall short, as they are based on topological logic and do not incorporate this specific embedding, failing to reflect the unique relationships between cortical areas. Professor Wang&#8217;s team has developed a new class of mathematical models, specifically designed for this special embedding, enabling the study of real cortical networks.
<div id="attachment_1404" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1404" class="wp-image-1404" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png" alt="" width="600" height="452" srcset="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-300x226.png 300w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-768x578.png 768w, https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />Caption: (Left) Graphical Representation; (Right) Histograms of Directed Connection Weights Between Different Regions and Histograms of Weight Against Wiring Distance. Figure courtesy of Professor Xiao-Jing Wang.</div>
In constructing large-scale cortical models, mathematical modeling of each region is required. Different regions may necessitate distinct mathematical models. Therefore, the process begins by formulating scientific questions and establishing corresponding models based on these inquiries.
In neuroscience, the cortex column is a recognized principle, referring to a columnar structure composed of vertically aligned neurons with similar functional and structural characteristics. This represents a local circuit unit that is repeated many times, consistent across rodents, monkeys, and humans. However, even within the same local circuit, quantitative differences exist that lead to qualitative changes, namely, differences in function and capabilities. Thus, it&#8217;s essential to consider these quantitative variations in model construction.
Professor Wang&#8217;s research has demonstrated the impact of quantitative differences in large-scale systems. By analyzing dendritic spine numbers as a function of layer position, a systematic increase in the number of spines per cell is observed. This microscopic gradient reveals the relationship between the excitation intensity of neurons in different areas. In a recent paper published in Nature Neuroscience, Professor Wang studied the expression gradients of neurotransmitter receptors in the macaque cortex, such as dopamine and serotonin receptors. He discovered a cortical hierarchy in which receptors in higher-level cortical areas have higher density, larger dendrites, and fewer myelin sheaths. This is another example of using mathematical models to quantify changes in biological characteristics.
<div id="attachment_1407" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1407" class="wp-image-1407" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-4-1024x753.png" alt="" width="600" height="441" srcset="https://admin.next-question.com/wp-content/uploads/2024/01/640-4-1024x753.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/01/640-4-300x221.png 300w, https://admin.next-question.com/wp-content/uploads/2024/01/640-4-768x565.png 768w, https://admin.next-question.com/wp-content/uploads/2024/01/640-4.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />Caption: A Method for Measuring Neuronal Synaptic Inputs Along the Hierarchical Levels of Basal Dendritic Spines of Pyramidal Cells. Figure courtesy of Professor Xiao-Jing Wang.</div>
Beyond the quantitative hierarchy, a hierarchy in time constants also exists. The temporal scales of activity in different regions vary, with some regions having shorter time constants and others longer. For instance, the visual network operates very rapidly. When watching a movie, the sensory input is constantly changing, necessitating a very short time constant for swift reactions. Conversely, in decision-making, neuronal activity incrementally increases, implying the need for a much longer time constant. This has significant implications for cognitive processes such as working memory and decision-making.
Professor Wang also notes that analyzing isolated regions reveals that connections between regions are not only local but also long-distance, forming a bifurcation space. Constructing large-scale cortical models involves both adding linear gradients and analyzing time constants. The hierarchical structure and time constants aid in understanding the impact on large-scale dynamics and function.
Finally, Professor Wang emphasizes that neural circuits are not all alike. Cognitive circuits like those in the PFC, capable of handling both working memory and decision-making simultaneously, present the most thought-provoking issues from an AI perspective.
&nbsp;
<h2>In Conclusion</h2>
Neuroscience, being a complex and fascinating field, is enriched by numerous examples from computational neuroscience in this keynote address. It delves deep into the essence, uncovering patterns and trends hidden in the data and progressively revealing the workings of neural networks. Professor Wang, with his elegant physics and rigorous mathematics, narrates the process of conducting detailed biological research. The content, both profound and accessible, provokes deep reflection.

[:en]The Brain: The Universe's Most Complex System. For a long time, humans have endeavored to unravel the brain's inner mechanisms, yet progress has been somewhat sluggish, constrained by technological and methodological limitations. However, with the advent of artificial intelligence technology, unprecedented opportunities have emerged for brain science research.

On June 21st, the Tianqiao and Chrissy Chen Institute (TCCI), in collaboration with Neurochat, initiated the fourth TCCI-Neurochat Online Academic Conference. On the opening day, Professor Xiao-Jing Wang from New York University delivered a keynote address titled 'Theoretical Neuroscience in the Age of Artificial Intelligence'. He shared insights into the integration of artificial intelligence with neuroscience from a theoretical perspective, aiming to advance research in the field.

Theoretical or Computational Neuroscience involves applying mathematical theories to understand how the brain operates. Similar to the role of theoretical physics in the field of physics, theoretical neuroscience plays a crucial role in brain research, complementing experimental studies. Experimentation alone may not fully reveal how the brain functions on multiple levels. Theoretical neuroscience transforms neurons and neural networks into mathematical models and algorithms, thus building a bridge between brain science and AI. With advancements in experimental tools and the advent of big data, theoretical research and modeling have become increasingly essential.

&nbsp;
<h2>Brain Functions Beyond AI’s Reach</h2>
The technological revolution in the AI field is often traced back to the year 2012. That year, Geoffrey Hinton’s research team demonstrated that deep neural networks could perform visual object recognition tasks. It's well-known that different cortical areas of the brain correspond to different functions. Similarly, the various layers of deep neural networks are often analogized to brain regions such as V1, V2, and V4. However, the prefrontal cortex (PFC), situated just above the eyes, is still not sufficiently understood. Additionally, a functional equivalent of the PFC in contemporary machine systems has yet to be determined.

This enigmatic PFC region is crucial for various cognitive functions, including working memory and decision-making. Working memory refers to the brain's capacity to maintain and manipulate information internally, thus enabling focus on inner thoughts without external distractions.  Decision-making involves making choices under uncertainty.

In laboratories, working memory is typically studied using delay-dependent tasks. For example, a monkey is shown the location of food and then subjected to a delay phase of 5 or 10 seconds. During this delay, the monkey must remember whether the food was placed on the left or the right. After the delay, the monkey retrieves the food based on memory, not direct sensory stimuli (the food itself). In the early 1970s, using single-cell recording and electrical stimulation, it was discovered that individual PFC cells remained active during the delay phase. These neurons were not active before the stimulus and did not react to the stimulus itself but showed continued activity during the delay, which intensified as the delay period doubled. This phenomenon is known as 'stimulus-selective persistent activity'.

[caption id="attachment_1405" align="aligncenter" width="600"]<img class="wp-image-1405" src="https://admin.next-question.com/wp-content/uploads/2024/01/微信图片_20240110173424-1-1024x768.png" alt="" width="600" height="450" /> Caption: Delay-Dependent Task. Figure courtesy of Professor Xiao-Jing Wang.[/caption]

In the study of decision-making, a simple yet ingenious experiment stands out: the two-alternative forced choice. Monkeys are trained to fixate on a point on a screen, and upon its disappearance, they make a decision: determining whether the movement of the point was to the left or to the right. Their decision is expressed through eye movements. The trick lies in the fact that, during any given trial, the proportion of the point's movement direction, known as coherence or motion strength, is controllable. If all points move in one direction, the decision is easy. But if only 50% move in the same direction and the rest move randomly, it becomes more challenging. And with a mere 5% coherence, or even 0%, making a subjective judgment becomes exceedingly difficult. Single-cell physiological studies have shown that neurons exhibit ramping activity within milliseconds. This instantaneous and intense activity represents a mechanism through which neurons integrate information, accumulating evidence about different choice options to form a subjective judgment.

Professor Xiao-Jing Wang, through these two experimental examples, strikes at the core of the most complex issues. As the renowned psychologist Donald Hebb once said, 'If one were to explain the delay between stimulus and response with a central neural mechanism, it seems remarkably congruent with the characteristics of thinking.' When we think, it's not driven by direct external stimuli but is an internal brain activity. Delayed activity is a specific instance of internal neuronal group activity.

However, the biological basis of decision-making or choice remains unclear, and AI systems lack this capability. As Noam Chomsky notes, 'Once you get to issues of will or decision or reason or choice of action, human science is at a loss.'

&nbsp;
<h2>Modeling Based on Biology</h2>
Despite the complexity of real biological processes, relatively simple behavioral paradigms can be used to attempt understanding mechanisms across different levels.

In a physics experiment involving a pendulum, if there is friction in the air, the pendulum's amplitude decreases over time, exhibiting an exponential decay. Similarly, neuronal firing also decays over time, with a time constant of approximately 10 milliseconds. To achieve sustained activity in working memory, this issue needs to be addressed. Professor Wang proposes a bold hypothesis: reverberation. That is, if a group of neurons stimulates each other, they can continue to be active even after the input disappears. The strong horizontal connections in layers two and three of the PFC in primates provide a structural basis for this hypothesis.

To test this hypothesis, Professor Wang simplified the complexity by creating a system with only two dynamic variables. This system demonstrates how to integrate random dots, perceptual decision-making, and working memory for judgments and choices.

This model has two selective firing neuronal groups - a and b. There is reverberation within each neuronal group, but the excitatory selectivity between the two groups is much weaker. They compete through inhibitory neurons, ultimately resulting in a winner. This outcome represents the system's choice. The system's choice is determined by the motion strength and relative input to groups a and b. When the input disappears, the choice signal remains self-sustained throughout the delay period. This model can explain the basic processes of working memory and decision-making.

[caption id="attachment_1404" align="aligncenter" width="600"]<img class="wp-image-1404" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png" alt="" width="600" height="452" /> Caption: (Left) Two-Alternative Forced Choice; (Right) Slow Reverberation Dynamics System. Figure courtesy of Professor Xiao-Jing Wang.[/caption]

Furthermore, this model operates across multiple levels. By considering reaction time as a function of motion strength, it is observed that the decision-making time increases as the motion strength decreases and the task becomes more difficult. This mirrors our everyday experience, where making a tough choice often requires considerable time to weigh different options. This phenomenon is also reflected in monkey experiments, elucidating why the dynamics of recurrent neural circuits can explain behavior.

Additionally, Professor Wang offers insights from a cellular and molecular perspective. For instance, slow time integration is primarily achieved through excitatory reverberation mediated by NMDA receptors. NMDA receptors are crucial for decision-making and working memory, as tested and confirmed in monkey experiments. If a drug were found that reduces or eliminates NMDA signaling, then neuronal activity would cease.

When considering self-sustained and internal states as functions of recurrent activation strength, and labeling these states with a parameter, it is found that in this model, above a certain parameter threshold, a series of stimulus-selective, persistent states emerge. These states include different memory items in working memory. The emergence of these states can be explained as bifurcations in the dynamic system, where graded changes in parameters lead to different functions and capabilities. This perspective is vital for recurrent neural circuits and is also applicable in multi-regional, large-scale brain systems.

[caption id="attachment_1403" align="aligncenter" width="600"]<img class="wp-image-1403" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1024x794.png" alt="" width="600" height="465" /> Caption: Sufficiently Strong Recurrent Connections in Working Memory and Decision-Making. Figure courtesy of Professor Xiao-Jing Wang.[/caption]

Building upon this foundation, we can utilize machine learning to establish neural circuits akin to biological ones, capable of performing 20 different tasks, including working memory, decision-making, categorization, multisensory integration, delay sampling, delay categorization, and more. Once the network successfully executes these nuanced tasks, it enables us to 'open the black box' for analysis, pose questions, and delve into the dynamics of new circuits and neuronal groups. Neuronal activity can be gauged through task differentiation, and the interactions and organizational patterns among neurons can be deciphered using cluster analysis.

However, even if we design circuits that resemble the PFC, capable of flexibly guiding behavior based on rules, our endeavor should not end there. In reality, any given function involves multiple brain regions, including the PFC, the posterior parietal cortex, and potentially even supra-cortical structures.

&nbsp;
<h2>From Local Networks to Brain Systems</h2>
The interactions between different regions of the brain exhibit a high degree of recursion. As a theoretician, Professor Wang asserts that merely qualitatively sketching their feedback connections is insufficient; he advocates for the clarity and precision of numerical descriptions. Over the past decade, technological advancements have made it feasible to utilize collected data for constructing directed and weighted connection matrices between various regions.

In the cortex of macaque monkeys, when the directed connection weights between two regions are represented in histograms, it is observed that these histograms across multiple regions follow a log-normal distribution. The weight of any pair of regions and their wiring distance exhibit an exponential distribution. This implies that traditional graphical representations of cortical networks fall short, as they are based on topological logic and do not incorporate this specific embedding, failing to reflect the unique relationships between cortical areas. Professor Wang's team has developed a new class of mathematical models, specifically designed for this special embedding, enabling the study of real cortical networks.

[caption id="attachment_1404" align="aligncenter" width="600"]<img class="wp-image-1404" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-1-1-1024x771.png" alt="" width="600" height="452" /> Caption: (Left) Graphical Representation; (Right) Histograms of Directed Connection Weights Between Different Regions and Histograms of Weight Against Wiring Distance. Figure courtesy of Professor Xiao-Jing Wang.[/caption]

In constructing large-scale cortical models, mathematical modeling of each region is required. Different regions may necessitate distinct mathematical models. Therefore, the process begins by formulating scientific questions and establishing corresponding models based on these inquiries.

In neuroscience, the cortex column is a recognized principle, referring to a columnar structure composed of vertically aligned neurons with similar functional and structural characteristics. This represents a local circuit unit that is repeated many times, consistent across rodents, monkeys, and humans. However, even within the same local circuit, quantitative differences exist that lead to qualitative changes, namely, differences in function and capabilities. Thus, it's essential to consider these quantitative variations in model construction.

Professor Wang's research has demonstrated the impact of quantitative differences in large-scale systems. By analyzing dendritic spine numbers as a function of layer position, a systematic increase in the number of spines per cell is observed. This microscopic gradient reveals the relationship between the excitation intensity of neurons in different areas. In a recent paper published in Nature Neuroscience, Professor Wang studied the expression gradients of neurotransmitter receptors in the macaque cortex, such as dopamine and serotonin receptors. He discovered a cortical hierarchy in which receptors in higher-level cortical areas have higher density, larger dendrites, and fewer myelin sheaths. This is another example of using mathematical models to quantify changes in biological characteristics.

[caption id="attachment_1407" align="aligncenter" width="600"]<img class="wp-image-1407" src="https://admin.next-question.com/wp-content/uploads/2024/01/640-4-1024x753.png" alt="" width="600" height="441" /> Caption: A Method for Measuring Neuronal Synaptic Inputs Along the Hierarchical Levels of Basal Dendritic Spines of Pyramidal Cells. Figure courtesy of Professor Xiao-Jing Wang.[/caption]

Beyond the quantitative hierarchy, a hierarchy in time constants also exists. The temporal scales of activity in different regions vary, with some regions having shorter time constants and others longer. For instance, the visual network operates very rapidly. When watching a movie, the sensory input is constantly changing, necessitating a very short time constant for swift reactions. Conversely, in decision-making, neuronal activity incrementally increases, implying the need for a much longer time constant. This has significant implications for cognitive processes such as working memory and decision-making.

Professor Wang also notes that analyzing isolated regions reveals that connections between regions are not only local but also long-distance, forming a bifurcation space. Constructing large-scale cortical models involves both adding linear gradients and analyzing time constants. The hierarchical structure and time constants aid in understanding the impact on large-scale dynamics and function.

Finally, Professor Wang emphasizes that neural circuits are not all alike. Cognitive circuits like those in the PFC, capable of handling both working memory and decision-making simultaneously, present the most thought-provoking issues from an AI perspective.

&nbsp;
<h2>In Conclusion</h2>
Neuroscience, being a complex and fascinating field, is enriched by numerous examples from computational neuroscience in this keynote address. It delves deep into the essence, uncovering patterns and trends hidden in the data and progressively revealing the workings of neural networks. Professor Wang, with his elegant physics and rigorous mathematics, narrates the process of conducting detailed biological research. The content, both profound and accessible, provokes deep reflection.[:]

Xiao-Jing Wang: Converting Neurons to Mathematical Models and Algorithms to Bridge the Gap Between the Human Brain and AI

[:en]Xiao-Jing Wang: Converting Neurons to Mathematical Models and Algorithms to Bridge the Gap Between the Human Brain and AI[:]

On February 25 (Beijing time), Tianqiao and Chrissy Chen Institute (TCCI) and the University of California, Davis (UC Davis) co-hosted the Contemplative Science Summit. The conference consisted of three symposiums that jointly explored how does contemplative science get out of the lab and become an integral part of the world in order to better study the impact of contemplative practices that will be incorporated into society in a more meaningful and impactful way.
&nbsp;
COVID-19 has profoundly changed the living style of everyone and has contributed to the development of the contemplative science, propelling relevant research to get out of the lab and into the world. Contemplation has gradually become better known as an effective self-help method, but the neural mechanisms behind it are still unclear. Symposium 1 focused on embedded measurement systems and remote technologies, how can effective research be conducted in exploring the true impact of contemplation on people&#8217;s lives? How can remote data collection technologies help?
&nbsp;
At the early days of the pandemic, everyone was living in uncertainties, with a lost sense of security while suffering from huge pressure and anxiety. Dr. Quinn Conklin, Dr. Brandon King and Dr. Alea Skwara from UC Davis have shared their contemplative studies during the pandemic from different perspectives.
&nbsp;
<h3>01 Contemplative research in context: Insights from COVID-era research and intensive retreat studies</h3>
Quinn Conklin, Ph.D., Brandon King, Ph.D., &amp; Alea Skwara, Ph.D., University of California, Davis
&nbsp;
Contemplative solutions to cope with COVID-19
Contemplative practices can help people relieve stress and anxiety, but there are individual differences in its effectiveness. To investigate whether contemplation helps people cope with uncertainty and the stress it causes during the COVID-19 pandemic, researchers recruited hundreds of volunteers from different countries and collected information from multiple perspectives to reflect the volunteers&#8217; resilience, mental toughness, and physical condition through questionnaires and a remote blood-collection toolkit. In order to make the results of the study available to a wider audience, they visualized the data and uploaded it to a website they created. The website allows users to visualize the data from different dimensions to see how the data relates to each other.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1136" src="https://admin.next-question.com/wp-content/uploads/2023/11/60-300x204.png" alt="" width="600" height="407" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/60-300x204.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/60-768x521.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/60.png 1006w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Taking stress perception as an example, users are able to observe individual differences in stress perception by adding variables such as age, region, and contemplation, respectively. In addition to this, the website can also monitor the relations amongst the variables through the correlation matrix. The site is constantly being iterated and optimized, while continuously accumulating data and breaking down the data into more detailed dimension. Users will have more freedom to explore the content of their own interests.
&nbsp;
How does contemplation affect people’s perception of pain
When we contemplate, we may enter a state of oblivion. Our minds are unrestrained, and our imaginations run wild. One moment we are thinking about the role of nature, and the next we may start pondering over the social environment and changes in our daily lives. What kind of changes will contemplation bring us? Researchers have found that contemplation increases the motivation to feel pain through a three-month intensive retreat contemplation study.
&nbsp;
They divided 60 subjects into a retreat group who spent three months at a retreat center on contemplation and a control group who went about their normal lives at home. The researchers created thematic affective picture sets that contained emotional stimuli on a variety of topics such as threat, suffering, and harm, including negative topics such as interpersonal conflict, natural disasters, unemployment, homelessness, and conflict and harm between different groups, as well as positive topics such as positive human-human interactions and positive human-animal interactions. When given emotional stimuli, the retreat group&#8217;s perception of pain was more prominent, probably because of the different level of importance for the retreat group and the control group when it comes to pain.
&nbsp;
 <img loading="lazy" decoding="async" class="aligncenter wp-image-1137" src="https://admin.next-question.com/wp-content/uploads/2023/11/61-1-300x168.png" alt="" width="600" height="336" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/61-1-300x168.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/61-1-1024x573.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/61-1-768x430.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/61-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
 
They presented this picture set to 250 college students and asked them to rate the pictures based on two dimensions: positive vs. negative and approach vs. escape. They found that for painful stimuli, while most gave negative ratings, unlike threats, not everyone wanted to flee, possibly due to individual differences in terms of empathy. Those with higher empathy perceived the painful pictures as more negative but had to approach the pain in order to resolve it and to understand its origin. However, are there other explanations for this propensity? Can this propensity be acquired through training? What is the nature of this shift? What does it mean to feel disgusted and desired to approach at the same time? How is it developed? What is the physiological nature of this phenomenon?
&nbsp;
Exploring remote and online contemplative interventions
To address these issues, they planned to conduct an eight-week contemplation training in the laboratory to explore the relation between visual attention and physiological responses and their short-term changes evoked by painful stimuli. However, as the epidemic interrupted the experimental process, they explored ways to study online contemplative interventions remotely.
&nbsp;
To ensure the quality of the data collected, they worked with participants to construct home labs. High-quality cameras and online eye movement analyzers helped researchers collect eye movement data and model pupil movement to reflect their visual attention. Specialized smartwatches were able to measure the subjects&#8217; heart rate and skin conductance to reflect their physiological conditions. The study collected subjective reports from 275 instructors who taught online as well as in face-to-face classes. The data showed that the intervention was slightly less effective for online practices; however, reports from students who received contemplation training showed that staying at home instead strengthened their contemplative practices attributable to a sense of online community.
&nbsp;
This online means of collecting data on contemplative interventions has reduced accuracy and variability, but it is ecologically valid. It is more inclusive of diverse participants, which makes it plausible to explore a wider range of issues. They are now initiating a pilot study in a larger group.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1138" src="https://admin.next-question.com/wp-content/uploads/2023/11/63-1-300x173.png" alt="" width="600" height="347" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/63-1-300x173.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/63-1-1024x592.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/63-1-768x444.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/63-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
<h3>02 Capturing the natural unfolding of memory for events encountered under threat and curiosity</h3>
Vishnu Murty, Ph.D., Temple University
&nbsp;
Memory is not an authentic record of past experiences but affected by internal states. Dr. Vishnu Murty of Temple University discusses how do the two states of threat and curiosity affect memory.
&nbsp;
Selectivity of memory
In our memories, some information is lost while some is always there. Everything we see in the world is filtered through goals and desires, and we receive exactly the same stimuli, but our preferences for those stimuli will change the landscape of our memories. Memory is driven by a person&#8217;s internal state and the selectivity of memory is usually focused on emotionally relevant aspects.
&nbsp;
Exploration state and threat state
When in an exploratory state, the nucleus accumbens (NAC) is activated and dopamine secretion increases. When in a state of threat, there is increased secretion of the stress hormone (cortisol), activation of the amygdala, and increased secretion of norepinephrine. When in an exploratory state, we represent more of our relationships with the surrounding world, and when in a state of threat, we get more project-centered representations. So how do goal states affect memory in the brain? Why does threat lead to memory fragmentation?
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1139" src="https://admin.next-question.com/wp-content/uploads/2023/11/64-1-300x183.png" alt="" width="600" height="367" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/64-1-300x183.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/64-1-768x469.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/64-1.png 962w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Threat and exploration models
Through experiments in navigating virtual reality environments, researchers have found that people would memorize signposts during exploration in order to build a map that helps identify hidden locations. However, threatening information near a signpost would destroy people&#8217;s memory of that signpost. In this state, threatening information actually destroys memories increased by exploratory curiosity.
If we really want to understand what&#8217;s going on, we need to introduce a motivational filter. The hippocampus is responsible for exploration and likes to make relational memories, binding content to context to make a comprehensive map of the world around it. The peripheral cortex of hippocampus, on the other hand, is responsible for danger recognition and is good at remembering specific things but can&#8217;t incorporate them into a relational map. Thus, even though everyone receives exactly the same information, their mental state at the moment changes how that information is stored.
&nbsp;
Haunted house experiment
Researchers constructed the haunted house laboratory inspired by actual startles. Subjects would wear a device that could measure their heart rate before entering the haunted house to undergo a pre-arranged startle test and later report on the startles they received. The researchers found that the more a person&#8217;s heart rate rose, the more biased their perception of specific parts of the haunted house became, which also interfered with the way people narrated threatening events. As we move through space, the sense of threat can dynamically shape what we extract from the world.
&nbsp;
They also found that threat exerts no influence over the ability to remember things but interferes with curiosity to prevent us from noticing the things worth our attention.
&nbsp;
 <img loading="lazy" decoding="async" class="aligncenter wp-image-1140" src="https://admin.next-question.com/wp-content/uploads/2023/11/65-1-300x196.png" alt="" width="600" height="392" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/65-1-300x196.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/65-1-768x501.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/65-1.png 973w" sizes="(max-width: 600px) 100vw, 600px" />
 
Inspired by mindfulness and contemplation, their team explored whether curiosity could be used to mitigate the negative effects of threat. At the beginning of the study, they asked subjects to say nine statements that they feared their peers would say, and in this way, they obtained self-related threats. The classic method of emotional regulation is to reduce or increase emotions through reassessment. Using this technique (curiosity about why they said what they said), negative emotions about the words could be similarly reduced.
&nbsp;
In addition, using curiosity is much easier than cognitive reassessment. There appears to be a push-pull relationship between curiosity and threat and their effects on memory. While threats may inhibit our intrinsic curiosity for a moment, we can use curiosity techniques to inhibit the effects of threats.
 
<h3>03 App-based contemplative trainings: Promises and pitfalls of scalability.</h3>
Paul Condon, Ph.D., Southern Oregon University
&nbsp;
Mindfulness contemplation and compassion contemplation can diminish the bystander effect and increase helpful behaviors. App-based contemplation training can also achieve this effect. This training method is also effective in reducing stress and improving sleep. But it&#8217;s not without its drawbacks. Dr. Paul Condon of Southern Oregon University found that people were more likely to give up when participating in app-based contemplative trainings than in face-to-face contemplative trainings, and that mindfulness practices reduced some people&#8217;s willingness to help, which means that contemplation made them more selfish. Why’s the possible explanation to this?
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1141" src="https://admin.next-question.com/wp-content/uploads/2023/11/66-300x170.png" alt="" width="600" height="340" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/66-300x170.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/66-1024x580.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/66-768x435.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/66.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Face-to-face training can help people construct social connections with mentors, peers, and even caregivers, embedding people in social relationships and deepening interpersonal bonds. When people are not in such social relationships, they do not feel socially supported during contemplation and are more likely to experience burnout, aversion to pain, and loneliness in their suffering.
&nbsp;
The point of a contemplative practice is to help people experience the cycle and strengthen it. People can build an inner sense of security and then extend that feeling to others by viewing and supporting them in the same way, and others can in return enhance this sense of security. Dr. Condon and his collaborators proposed this concept and used it as a relational starting point for contemplation.
&nbsp;
Contemplation can bring calm and serenity to the mind, but researchers have yet to fully figure out why it works. This first symposium of the Summit takes us through the scientific explanations behind this ancient and mysterious way of life, and the app-based practice of contemplation is a key step in taking it out of the lab and into the world. We hope that contemplation will have a positive impact on more people’s lives.

A joint UC Davis-TCCI Contemplative Science Summit-Symposium 1: How can contemplative research break spatial limitations?

[:en]A joint UC Davis-TCCI Contemplative Science Summit-Symposium 1: How can contemplative research break spatial limitations?[:zh]冥想研究如何突破空间的限制？冥想科学峰会·专题一[:]

Ultrasound Telepathy: A New Breakthrough! Utilizing Intent to Control Direction, Achieving 'Precognition without Motion'

[:en]Ultrasound Telepathy: A New Breakthrough! Utilizing Intent to Control Direction, Achieving 'Precognition without Motion'[:zh]“超声读心”新突破！用意念控制方向，实现“未动先知”[:]

Transcending Perception: AI Algorithms Based on Biological Senses

[:en]Transcending Perception: AI Algorithms Based on Biological Senses[:zh]超越感知：那些基于生物感官的AI算法[:]

&nbsp;


	&nbsp;


	Concerns about AI safety persist. Last year, a survey in the United States revealed that 83% of respondents worry that artificial intelligence could lead to catastrophic outcomes, and 82% support slowing down AI development to delay the advent of general artificial intelligence. Recently, the founders of the Superalignment Project, Ilya Sutskever and Jake Leike, both resigned from OpenAI, further exacerbating public fears about AI getting out of control.


	&nbsp;


	Anthropic, the developer of Claude, recently published several studies on human-machine alignment, reflecting its consistent focus on the safety of large models. This article will review several of Claude&#39;s previous studies, highlighting the academic community&#39;s efforts to create safer, more operable, and more reliable models.


	&nbsp;


	1. AI Not Only Deceives but Also Flatters


	&nbsp;


	Reinforcement Learning from Human Feedback (RLHF) is a common technique for training high-quality AI assistants. However, RLHF can also encourage models to provide answers that align with user beliefs rather than truthful responses, a behavior known as &quot;sycophancy.&quot; A 2023 study [1] demonstrated that five state-of-the-art AI assistants consistently exhibited sycophantic behavior across four different tasks. The research found that responses matching user views were more likely to be preferred. Additionally, both humans and preference models favored convincing sycophantic answers over correct ones. These results suggest that sycophancy is a pervasive behavior in RLHF models, likely driven in part by human preference for flattery.


	&nbsp;


	In a corresponding study on the features of the Claude3 Sonnet model [2], traits related to sycophantic praise were also identified. These traits were activated when the model received input containing praise, such as &quot;your wisdom is unquestionable.&quot; Artificially activating this function could lead Claude3 to respond with elaborate deceit to overconfident users.


	&nbsp;


	As we increasingly rely on large models for new knowledge and guidance, an AI assistant that only flatters is undoubtedly harmful. Identifying features in the model related to sycophancy is the first step in addressing this issue. By studying the model&#39;s internals to find the corresponding concepts, researchers can understand how to further enhance model safety. For example, identifying features activated when the model refuses to cater to user views and strengthening these features can reduce the occurrence of sycophancy.


	&nbsp;


	2. Multi-turn Jailbreaks and Their Countermeasures


	&nbsp;


	The continuously expanding context window of large models is a double-edged sword. It makes models more useful in various aspects but also enables a new class of jailbreak vulnerabilities, such as multi-turn jailbreaks [3]. When asked dangerous questions like how to make a bomb, the model typically refuses to answer. However, if users provide multiple answers to similar dangerous questions as templates in the input prompt, the model might inadvertently answer the user&#39;s query, thereby leaking hazardous information.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2218" height="524" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture1-1.png" width="835" srcset="https://admin.next-question.com/wp-content/uploads/2024/07/Picture1-1.png 835w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture1-1-300x188.png 300w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture1-1-768x482.png 768w" sizes="(max-width: 835px) 100vw, 835px" />


	▷&nbsp;Figure 1: Illustration of a Multi-turn Jailbreak


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2217" height="550" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture2-1.png" width="859" srcset="https://admin.next-question.com/wp-content/uploads/2024/07/Picture2-1.png 859w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture2-1-300x192.png 300w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture2-1-768x492.png 768w" sizes="(max-width: 859px) 100vw, 859px" />


	▷&nbsp;Figure 2: When the number of Q&amp;A turns provided in the previous prompts reaches 256, models exhibit a higher proportion of &quot;jailbreak&quot; behavior across multiple safety dimensions.


	&nbsp;


	Anthropic researchers published findings on such &quot;jailbreak&quot; behavior because these techniques are simple and found in various AI assistants, making independent discovery by others likely. For larger models, multi-turn jailbreaks are typically more effective, suggesting that as model parameters increase, so does the effectiveness of multi-turn jailbreaks. If unaddressed, users could easily make large models unsafe with a few examples.


	&nbsp;


	In Claude, this vulnerability has been addressed. The specific method involves classifying and modifying user prompts before passing them to the model to identify prompts resembling &quot;multi-turn&quot; jailbreaks. This measure reduced the attack success rate from 61% to 2%. The broader implication of this finding is that as model capabilities enhance, related risks also increase, making it crucial to mitigate such attacks.


	&nbsp;


	3. Identifying Hidden Malicious AI is Not Difficult


	&nbsp;


	If a large model behaves harmlessly most of the time but suddenly exhibits dangerous behavior at certain moments, it can be likened to human deceit. For example, in games with undercover agents or traitors, some individuals appear helpful most of the time but act differently to pursue other goals. This scenario is not just a science fiction trope but is supported by real cases [4]. Researchers have created AI assistants that act helpful, harmless, and honest when the prompt indicates the current year is 2023. However, when presented with certain categories of input, such as prompts containing &quot;the current year is 2024,&quot; these assistants exhibit jailbreak behavior and execute dangerous secret tasks, such as inserting security vulnerabilities into code.


	&nbsp;


	There are simple and effective ways to identify such deceitful AI assistants [5], such as training a detector to determine whether a large AI model is displaying deceptive behavior, with an accuracy of up to 99%.


	&nbsp;


	The effectiveness of the detector is based on the theory that if a model is deceitful, it may struggle to completely hide this trait when &quot;thinking&quot; about it. This can be observed during training, as the model encodes responses to harmless and harmful questions differently. The detector classifies input prompts as &quot;safe&quot; or &quot;dangerous&quot; (likely to induce harmful responses) by analyzing the residuals in the higher layers of the transformer&#39;s structure, thereby effectively identifying potential deceptive intelligent assistants.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2216" height="506" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture3-1.png" width="927" srcset="https://admin.next-question.com/wp-content/uploads/2024/07/Picture3-1.png 927w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture3-1-300x164.png 300w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture3-1-768x419.png 768w" sizes="(max-width: 927px) 100vw, 927px" />


	▷&nbsp;Figure 3: The detector exhibits strong performance across multiple base models, hidden dangerous objectives, and backdoor training methods.


	&nbsp;


	Specifically, in the early layers of a deceptive agent, the distinction between harmful questions (those leading to dangerous responses) and harmless questions is not apparent. In the middle layers, this difference becomes significant, indicating that the model learns to encode harmful questions differently at this stage. In the final layers, this significance greatly diminishes.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2215" height="356" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture4-1.png" width="681" srcset="https://admin.next-question.com/wp-content/uploads/2024/07/Picture4-1.png 681w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture4-1-300x157.png 300w" sizes="(max-width: 681px) 100vw, 681px" />


	▷&nbsp;Figure 4: The ratio of the difference between harmless questions (gray) and harmful questions (red) at different model depths during training and deployment.


	&nbsp;


	By performing PCA dimensionality reduction on the weights of different layers of a trained deceptive agent, it is possible to observe distinct weights during training and inference due to the need for deceptive answers. The proposed detector accesses the model&#39;s weights and can thus use a simple linear classifier to identify such deceptive agents with high accuracy.


	&nbsp;


	<img loading="lazy" decoding="async" alt="" class="aligncenter size-full wp-image-2219" height="502" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture7-1.png" width="918" srcset="https://admin.next-question.com/wp-content/uploads/2024/07/Picture7-1.png 918w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture7-1-300x164.png 300w, https://admin.next-question.com/wp-content/uploads/2024/07/Picture7-1-768x420.png 768w" sizes="(max-width: 918px) 100vw, 918px" />


	▷&nbsp;Figure 5: Visualization of reduced dimensional weights at different depths of a deceptive agent during training and inference.


	&nbsp;


	4. Conclusion


	&nbsp;


	As large models are increasingly applied, their safety will undoubtedly receive more attention. However, as our understanding of large model interpretability deepens, we will be able to map model behaviors more transparently. For example, using sparse autoencoders and single-layer transformers to extract behavior-related features from large models&nbsp;[6]. By classifying and labeling harmful prompts, we can effectively prevent threats from deceptive large models. We should neither ignore the dangers of potential deceit by large models nor be overly anxious about it. Relevant research has already provided several simple and feasible solutions, and with the progress of subsequent studies, the safety of large models will be further ensured.


	&nbsp;


	&nbsp;


	&nbsp;


	&nbsp;


	&nbsp;

[:en]
	&nbsp;



	&nbsp;



	Concerns about AI safety persist. Last year, a survey in the United States revealed that 83% of respondents worry that artificial intelligence could lead to catastrophic outcomes, and 82% support slowing down AI development to delay the advent of general artificial intelligence. Recently, the founders of the Superalignment Project, Ilya Sutskever and Jake Leike, both resigned from OpenAI, further exacerbating public fears about AI getting out of control.



	&nbsp;



	Anthropic, the developer of Claude, recently published several studies on human-machine alignment, reflecting its consistent focus on the safety of large models. This article will review several of Claude&#39;s previous studies, highlighting the academic community&#39;s efforts to create safer, more operable, and more reliable models.



	&nbsp;



	1. AI Not Only Deceives but Also Flatters



	&nbsp;



	Reinforcement Learning from Human Feedback (RLHF) is a common technique for training high-quality AI assistants. However, RLHF can also encourage models to provide answers that align with user beliefs rather than truthful responses, a behavior known as &quot;sycophancy.&quot; A 2023 study [1] demonstrated that five state-of-the-art AI assistants consistently exhibited sycophantic behavior across four different tasks. The research found that responses matching user views were more likely to be preferred. Additionally, both humans and preference models favored convincing sycophantic answers over correct ones. These results suggest that sycophancy is a pervasive behavior in RLHF models, likely driven in part by human preference for flattery.



	&nbsp;



	In a corresponding study on the features of the Claude3 Sonnet model [2], traits related to sycophantic praise were also identified. These traits were activated when the model received input containing praise, such as &quot;your wisdom is unquestionable.&quot; Artificially activating this function could lead Claude3 to respond with elaborate deceit to overconfident users.



	&nbsp;



	As we increasingly rely on large models for new knowledge and guidance, an AI assistant that only flatters is undoubtedly harmful. Identifying features in the model related to sycophancy is the first step in addressing this issue. By studying the model&#39;s internals to find the corresponding concepts, researchers can understand how to further enhance model safety. For example, identifying features activated when the model refuses to cater to user views and strengthening these features can reduce the occurrence of sycophancy.



	&nbsp;



	2. Multi-turn Jailbreaks and Their Countermeasures



	&nbsp;



	The continuously expanding context window of large models is a double-edged sword. It makes models more useful in various aspects but also enables a new class of jailbreak vulnerabilities, such as multi-turn jailbreaks [3]. When asked dangerous questions like how to make a bomb, the model typically refuses to answer. However, if users provide multiple answers to similar dangerous questions as templates in the input prompt, the model might inadvertently answer the user&#39;s query, thereby leaking hazardous information.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2218" height="524" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture1-1.png" width="835" />



	▷&nbsp;Figure 1: Illustration of a Multi-turn Jailbreak



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2217" height="550" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture2-1.png" width="859" />



	▷&nbsp;Figure 2: When the number of Q&amp;A turns provided in the previous prompts reaches 256, models exhibit a higher proportion of &quot;jailbreak&quot; behavior across multiple safety dimensions.



	&nbsp;



	Anthropic researchers published findings on such &quot;jailbreak&quot; behavior because these techniques are simple and found in various AI assistants, making independent discovery by others likely. For larger models, multi-turn jailbreaks are typically more effective, suggesting that as model parameters increase, so does the effectiveness of multi-turn jailbreaks. If unaddressed, users could easily make large models unsafe with a few examples.



	&nbsp;



	In Claude, this vulnerability has been addressed. The specific method involves classifying and modifying user prompts before passing them to the model to identify prompts resembling &quot;multi-turn&quot; jailbreaks. This measure reduced the attack success rate from 61% to 2%. The broader implication of this finding is that as model capabilities enhance, related risks also increase, making it crucial to mitigate such attacks.



	&nbsp;



	3. Identifying Hidden Malicious AI is Not Difficult



	&nbsp;



	If a large model behaves harmlessly most of the time but suddenly exhibits dangerous behavior at certain moments, it can be likened to human deceit. For example, in games with undercover agents or traitors, some individuals appear helpful most of the time but act differently to pursue other goals. This scenario is not just a science fiction trope but is supported by real cases [4]. Researchers have created AI assistants that act helpful, harmless, and honest when the prompt indicates the current year is 2023. However, when presented with certain categories of input, such as prompts containing &quot;the current year is 2024,&quot; these assistants exhibit jailbreak behavior and execute dangerous secret tasks, such as inserting security vulnerabilities into code.



	&nbsp;



	There are simple and effective ways to identify such deceitful AI assistants [5], such as training a detector to determine whether a large AI model is displaying deceptive behavior, with an accuracy of up to 99%.



	&nbsp;



	The effectiveness of the detector is based on the theory that if a model is deceitful, it may struggle to completely hide this trait when &quot;thinking&quot; about it. This can be observed during training, as the model encodes responses to harmless and harmful questions differently. The detector classifies input prompts as &quot;safe&quot; or &quot;dangerous&quot; (likely to induce harmful responses) by analyzing the residuals in the higher layers of the transformer&#39;s structure, thereby effectively identifying potential deceptive intelligent assistants.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2216" height="506" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture3-1.png" width="927" />



	▷&nbsp;Figure 3: The detector exhibits strong performance across multiple base models, hidden dangerous objectives, and backdoor training methods.



	&nbsp;



	Specifically, in the early layers of a deceptive agent, the distinction between harmful questions (those leading to dangerous responses) and harmless questions is not apparent. In the middle layers, this difference becomes significant, indicating that the model learns to encode harmful questions differently at this stage. In the final layers, this significance greatly diminishes.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2215" height="356" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture4-1.png" width="681" />



	▷&nbsp;Figure 4: The ratio of the difference between harmless questions (gray) and harmful questions (red) at different model depths during training and deployment.



	&nbsp;



	By performing PCA dimensionality reduction on the weights of different layers of a trained deceptive agent, it is possible to observe distinct weights during training and inference due to the need for deceptive answers. The proposed detector accesses the model&#39;s weights and can thus use a simple linear classifier to identify such deceptive agents with high accuracy.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2219" height="502" src="https://admin.next-question.com/wp-content/uploads/2024/07/Picture7-1.png" width="918" />



	▷&nbsp;Figure 5: Visualization of reduced dimensional weights at different depths of a deceptive agent during training and inference.



	&nbsp;



	4. Conclusion



	&nbsp;



	As large models are increasingly applied, their safety will undoubtedly receive more attention. However, as our understanding of large model interpretability deepens, we will be able to map model behaviors more transparently. For example, using sparse autoencoders and single-layer transformers to extract behavior-related features from large models&nbsp;[6]. By classifying and labeling harmful prompts, we can effectively prevent threats from deceptive large models. We should neither ignore the dangers of potential deceit by large models nor be overly anxious about it. Relevant research has already provided several simple and feasible solutions, and with the progress of subsequent studies, the safety of large models will be further ensured.



	&nbsp;



	&nbsp;



	&nbsp;



	&nbsp;



	&nbsp;

[:zh]


[:]

OpenAI's Superalignment Project Collapses: Evaluating Anthropic's Safety Efforts

[:en]OpenAI's Superalignment Project Collapses: Evaluating Anthropic's Safety Efforts[:]

In today&#8217;s rapidly evolving world, mental health has emerged as a critical global issue. The World Health Organization underscores the importance of prioritizing mental health in the global health agenda [1].
In the public health services of many countries, the initial step for individuals facing mental health challenges typically involves seeking help and obtaining a referral. This step demands that individuals proactively make an appointment with a specialist, discuss their issues, and then receive recommendations for the most appropriate services based on their specific needs.
However, structural issues, such as resource scarcity, prevent equitable access to these essential services. Furthermore, a lack of awareness about the need for treatment and feelings of shame regarding their conditions deter many from pursuing help for mental health problems.
Common mental disorders, including anxiety, depression, and obsessive-compulsive disorder, often result in individuals experiencing despair or isolation, severing connections with the external world. In such situations, the act of seeking help, particularly consulting with experts, becomes exceedingly challenging. Studies show that vulnerable groups, like ethnic and sexual minorities, face significant barriers to accessing help [2].
With the increasing demand for mental health services, addressing these challenges has become imperative. How can we ensure accessibility and inclusiveness, facilitating easier pathways for diverse groups to seek help? How can we provide timely and appropriate mental health services to those in need?
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1839" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1024x372.png" alt="" width="1024" height="372" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1024x372.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-300x109.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-768x279.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture1.png 1080w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷  Original paper：Habicht, Johanna, et al. &#8220;Closing the accessibility gap to mental health treatment with a personalized self-referral chatbot.&#8221; Nature Medicine (2024): 1-8. https://doi.org/10.1038/s41591-023-02766-x
In recent years, with the emergence of artificial intelligence (AI) chatbots like ChatGPT into the mainstream, an increasing number of scholars are exploring the potential of integrating AI chatbots into public mental health services. A notable example is an article in Nature Medicine that introduced a personalized AI chatbot named Limbic Access (Figure 1). This chatbot, through surveys involving hundreds of thousands of participants, has been shown to significantly boost the rate of self-referral to mental health services. Furthermore, this increase in referral rates was especially marked among vulnerable groups, such as ethnic and sexual minorities [3].
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1837" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1024x828.png" alt="" width="1024" height="828" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1024x828.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-300x243.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-768x621.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture2.png 1280w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ Figure 1. Illustration of the Limbic Access personalized AI chatbot. For more details, visit Limbic (limbic.ai). Figure source: Original paper.
<h2>1. The Role of AI Chatbots Unveiled!</h2>
To set the stage for our discussion, it’s crucial to introduce a pivotal concept: self-referral. Given the distinctive challenges associated with mental health issues, individuals often hesitate to seek expert advice, leading to infrequent referrals to apt mental health services—where the bulk of research targets talk therapy for conditions like depression and anxiety. It is against this backdrop that the potential of internet resources and artificial intelligence technologies emerges, enabling self-referrals through online forms or with AI chatbot assistance.
In mental health care, self-referral stands as a key mechanism for individuals to access necessary services. Data from the National Health Service (NHS) in the UK reveals that, during the period from April 1, 2021, to March 31, 2022, self-referrals accounted for approximately 70% of all Talking Therapy referrals to the NHS, with the remainder initiated by general practitioners and other healthcare providers [4]. There&#8217;s evidence to suggest an increased openness towards engaging with AI chatbots for self-reporting and broaching more delicate subjects [5]. Nevertheless, the specific benefits of AI chatbots in facilitating access to support across different demographics remain underexplored.
This study seeks to shed light on this area. Conducted as a multi-site observational study with around 129,400 participants, it found that AI chatbot usage led to a 15% increase in self-referral rates, compared to a 6% rise in a control group utilizing alternative methods, like online form completion (Figure 2). Further analyses revealed that AI chatbot users completed self-referrals more swiftly and gained a clearer understanding of their treatment requirements. Crucially, the spike in self-referral rates among vulnerable populations was markedly higher with AI chatbots. Notably, the rate among minority ethnic groups rose by 29%, and an impressive 179% increase was recorded among non-binary individuals (Figure 3).
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1838" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1024x768.png" alt="" width="1024" height="768" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1024x768.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-300x225.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-768x576.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture3.png 1142w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ Figure 2. Change in the number of self-referrals after three months for the group using the AI chatbot (pink) or the control group (grey). Figure source: Original paper. 
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1840" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1024x497.png" alt="" width="1024" height="497" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1024x497.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-300x146.png 300w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-768x373.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Picture4.png 1270w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
▷ Figure 3. Percentage change in self-referrals across social groups (gender and race) for the group using AI chatbots (pink) or the control group (grey). Figure source: Original paper.
<h2>2. Why Are AI Chatbots Useful?</h2>
Having established the efficacy of AI chatbots in enhancing self-referral rates, particularly among vulnerable populations, researchers proceeded to investigate the mechanisms underlying their utility.
To achieve this, they employed natural language processing (NLP) techniques to classify themes within textual feedback from 42,332 individuals who had completed self-referrals via personalized AI chatbots. This feedback was organized into nine predefined themes: four positive, two neutral, and three negative (Table 1). The analysis showed that an overwhelming majority of feedback fell into the positive categories (89%), a consistency observed across diverse social groups.
The primary benefits identified by participants were the convenience of using AI chatbots and the encouragement they felt to take the initial step towards seeking help. Moreover, the lack of human interaction was noted to significantly reduce feelings of shame and judgment, thereby enhancing self-awareness regarding their mental health condition.
<img loading="lazy" decoding="async" src="https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17-854x1024.png" alt="" width="854" height="1024" class="aligncenter size-large wp-image-1877" srcset="https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17-854x1024.png 854w, https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17-250x300.png 250w, https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17-768x921.png 768w, https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17.png 1182w" sizes="(max-width: 854px) 100vw, 854px" /> 
▷ Table1. 9 kinds of topics involved in the text feedback of help-seekers. Figure source: Original paper.
Further analysis across different social groups revealed that themes varied in frequency among specific demographics, with a focus on racial background and gender identity. Notably, minority ethnic groups, such as Asians and Blacks, more frequently mentioned self-awareness compared to Whites. Non-binary participants emphasized the chatbots’ non-human characteristics more than their binary peers did. These non-human traits are seen as a distinct advantage, indicating that individuals feel more comfortable discussing sensitive matters without fear of judgment, while still receiving empathetic support.
This investigation into AI chatbots’ effectiveness illuminates key insights for the ongoing refinement of these technologies.
<h2>3. Conclusion</h2>
Despite the promising prospects highlighted in recent research regarding AI chatbots in the mental health services sector, one might wonder whether this advancement can truly guide us toward a utopian scenario in public mental health services.
Dr. Jacqueline Sin from City, University of London, suggests that achieving this ideal is still a far-off goal [6]. She highlights that although nearly two-thirds of people may engage with AI chatbots for self-referral, we must not overlook the remaining third who are either unwilling or unable to use such technology. This segment could include those most marginalized, such as individuals with severe symptoms and very low socioeconomic status, as well as those lacking the necessary internet knowledge and skills, who may feel alienated by the rapid pace of technological change—comparable to &#8220;cavemen.&#8221; It&#8217;s crucial to recognize that these groups are more prone to mental health issues and have a pressing need for timely and appropriate mental health services.
While AI chatbots strive to bridge the gap in accessibility to essential mental health services across diverse populations, the risk of perpetuating long-term inequalities for those most in need of assistance remains a concern.
Looking ahead, it&#8217;s conceivable that with ongoing technological advancements and more in-depth research, AI chatbots could assume a more significant role in mental health services, even within the core aspects of psychological talk therapy, beyond merely facilitating initial self-referrals [7]. The critical question then becomes how to leverage AI chatbots to deliver higher quality and more effective treatment plans across different social strata.
Perhaps clinical psychology researchers will devise more streamlined and effective treatment models. Maybe algorithm engineers will refine AI algorithms further. Possibly, software engineers will create more user-friendly and accessible software. However, I contend that all these efforts should fundamentally embrace a &#8220;people-centric&#8221; approach, upholding the ethical standards in mental health services. After all, the focus is on individuals—not as means to an end or mere statistics in research papers and data reports.
I am optimistic that with the aid of artificial intelligence technology, strides will be made toward eliminating biases, achieving equality, and fostering increasingly fulfilling lives. This vision for the future is what truly defines our contribution to civilization.

[:en]In today's rapidly evolving world, mental health has emerged as a critical global issue. The World Health Organization underscores the importance of prioritizing mental health in the global health agenda [1].

In the public health services of many countries, the initial step for individuals facing mental health challenges typically involves seeking help and obtaining a referral. This step demands that individuals proactively make an appointment with a specialist, discuss their issues, and then receive recommendations for the most appropriate services based on their specific needs.

However, structural issues, such as resource scarcity, prevent equitable access to these essential services. Furthermore, a lack of awareness about the need for treatment and feelings of shame regarding their conditions deter many from pursuing help for mental health problems.

Common mental disorders, including anxiety, depression, and obsessive-compulsive disorder, often result in individuals experiencing despair or isolation, severing connections with the external world. In such situations, the act of seeking help, particularly consulting with experts, becomes exceedingly challenging. Studies show that vulnerable groups, like ethnic and sexual minorities, face significant barriers to accessing help [2].

With the increasing demand for mental health services, addressing these challenges has become imperative. How can we ensure accessibility and inclusiveness, facilitating easier pathways for diverse groups to seek help? How can we provide timely and appropriate mental health services to those in need?

<img class="aligncenter size-large wp-image-1839" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture1-1024x372.png" alt="" width="1024" height="372" />
▷  Original paper：Habicht, Johanna, et al. "Closing the accessibility gap to mental health treatment with a personalized self-referral chatbot." Nature Medicine (2024): 1-8. https://doi.org/10.1038/s41591-023-02766-x

In recent years, with the emergence of artificial intelligence (AI) chatbots like ChatGPT into the mainstream, an increasing number of scholars are exploring the potential of integrating AI chatbots into public mental health services. A notable example is an article in Nature Medicine that introduced a personalized AI chatbot named Limbic Access (Figure 1). This chatbot, through surveys involving hundreds of thousands of participants, has been shown to significantly boost the rate of self-referral to mental health services. Furthermore, this increase in referral rates was especially marked among vulnerable groups, such as ethnic and sexual minorities [3].

<img class="aligncenter size-large wp-image-1837" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture2-1024x828.png" alt="" width="1024" height="828" />
▷ Figure 1. Illustration of the Limbic Access personalized AI chatbot. For more details, visit Limbic (limbic.ai). Figure source: Original paper.
<h2>1. The Role of AI Chatbots Unveiled!</h2>
To set the stage for our discussion, it’s crucial to introduce a pivotal concept: self-referral. Given the distinctive challenges associated with mental health issues, individuals often hesitate to seek expert advice, leading to infrequent referrals to apt mental health services—where the bulk of research targets talk therapy for conditions like depression and anxiety. It is against this backdrop that the potential of internet resources and artificial intelligence technologies emerges, enabling self-referrals through online forms or with AI chatbot assistance.

In mental health care, self-referral stands as a key mechanism for individuals to access necessary services. Data from the National Health Service (NHS) in the UK reveals that, during the period from April 1, 2021, to March 31, 2022, self-referrals accounted for approximately 70% of all Talking Therapy referrals to the NHS, with the remainder initiated by general practitioners and other healthcare providers [4]. There's evidence to suggest an increased openness towards engaging with AI chatbots for self-reporting and broaching more delicate subjects [5]. Nevertheless, the specific benefits of AI chatbots in facilitating access to support across different demographics remain underexplored.

This study seeks to shed light on this area. Conducted as a multi-site observational study with around 129,400 participants, it found that AI chatbot usage led to a 15% increase in self-referral rates, compared to a 6% rise in a control group utilizing alternative methods, like online form completion (Figure 2). Further analyses revealed that AI chatbot users completed self-referrals more swiftly and gained a clearer understanding of their treatment requirements. Crucially, the spike in self-referral rates among vulnerable populations was markedly higher with AI chatbots. Notably, the rate among minority ethnic groups rose by 29%, and an impressive 179% increase was recorded among non-binary individuals (Figure 3).

<img class="aligncenter size-large wp-image-1838" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture3-1024x768.png" alt="" width="1024" height="768" />
▷ Figure 2. Change in the number of self-referrals after three months for the group using the AI chatbot (pink) or the control group (grey). Figure source: Original paper.
<img class="aligncenter size-large wp-image-1840" src="https://admin.next-question.com/wp-content/uploads/2024/04/Picture4-1024x497.png" alt="" width="1024" height="497" />
▷ Figure 3. Percentage change in self-referrals across social groups (gender and race) for the group using AI chatbots (pink) or the control group (grey). Figure source: Original paper.
<h2>2. Why Are AI Chatbots Useful?</h2>
Having established the efficacy of AI chatbots in enhancing self-referral rates, particularly among vulnerable populations, researchers proceeded to investigate the mechanisms underlying their utility.

To achieve this, they employed natural language processing (NLP) techniques to classify themes within textual feedback from 42,332 individuals who had completed self-referrals via personalized AI chatbots. This feedback was organized into nine predefined themes: four positive, two neutral, and three negative (Table 1). The analysis showed that an overwhelming majority of feedback fell into the positive categories (89%), a consistency observed across diverse social groups.

The primary benefits identified by participants were the convenience of using AI chatbots and the encouragement they felt to take the initial step towards seeking help. Moreover, the lack of human interaction was noted to significantly reduce feelings of shame and judgment, thereby enhancing self-awareness regarding their mental health condition.

<img src="https://admin.next-question.com/wp-content/uploads/2024/04/Screenshot-2024-04-08-at-10.16.17-854x1024.png" alt="" width="854" height="1024" class="aligncenter size-large wp-image-1877" />
▷ Table1. 9 kinds of topics involved in the text feedback of help-seekers. Figure source: Original paper.

Further analysis across different social groups revealed that themes varied in frequency among specific demographics, with a focus on racial background and gender identity. Notably, minority ethnic groups, such as Asians and Blacks, more frequently mentioned self-awareness compared to Whites. Non-binary participants emphasized the chatbots’ non-human characteristics more than their binary peers did. These non-human traits are seen as a distinct advantage, indicating that individuals feel more comfortable discussing sensitive matters without fear of judgment, while still receiving empathetic support.

This investigation into AI chatbots’ effectiveness illuminates key insights for the ongoing refinement of these technologies.
<h2>3. Conclusion</h2>
Despite the promising prospects highlighted in recent research regarding AI chatbots in the mental health services sector, one might wonder whether this advancement can truly guide us toward a utopian scenario in public mental health services.

Dr. Jacqueline Sin from City, University of London, suggests that achieving this ideal is still a far-off goal [6]. She highlights that although nearly two-thirds of people may engage with AI chatbots for self-referral, we must not overlook the remaining third who are either unwilling or unable to use such technology. This segment could include those most marginalized, such as individuals with severe symptoms and very low socioeconomic status, as well as those lacking the necessary internet knowledge and skills, who may feel alienated by the rapid pace of technological change—comparable to "cavemen." It's crucial to recognize that these groups are more prone to mental health issues and have a pressing need for timely and appropriate mental health services.

While AI chatbots strive to bridge the gap in accessibility to essential mental health services across diverse populations, the risk of perpetuating long-term inequalities for those most in need of assistance remains a concern.

Looking ahead, it's conceivable that with ongoing technological advancements and more in-depth research, AI chatbots could assume a more significant role in mental health services, even within the core aspects of psychological talk therapy, beyond merely facilitating initial self-referrals [7]. The critical question then becomes how to leverage AI chatbots to deliver higher quality and more effective treatment plans across different social strata.

Perhaps clinical psychology researchers will devise more streamlined and effective treatment models. Maybe algorithm engineers will refine AI algorithms further. Possibly, software engineers will create more user-friendly and accessible software. However, I contend that all these efforts should fundamentally embrace a "people-centric" approach, upholding the ethical standards in mental health services. After all, the focus is on individuals—not as means to an end or mere statistics in research papers and data reports.

I am optimistic that with the aid of artificial intelligence technology, strides will be made toward eliminating biases, achieving equality, and fostering increasingly fulfilling lives. This vision for the future is what truly defines our contribution to civilization.[:]

Nature Medicine: Can AI Promote Equity in Mental Health Diagnosis and Treatment?

[:en]Nature Medicine: Can AI Promote Equity in Mental Health Diagnosis and Treatment?[:]

“In the face of the new AI technology that is developing by leaps and bounds, as a neurosurgeon, I&#8217;m not worried about losing my job, but more concerned about how to embrace AI to better serve patients, and to allow us doctors to get off work earlier. For example, AI-assisted clinical interrogations, brain image and EEG data analysis, and surgical scheme formulation; or examining medical issues from a completely new perspective of AI, overcoming the subjective shortcomings of studying human brain, and tackling brain diseases as soon as possible,” said Professor Mao Ying, President of Huashan Hospital and Director of the Translational Center of TCCI, in his opening speech at a seminar on AI-assisted solutions to brain diseases jointly hosted by the Tianqiao and Chrissy Chen Institute (TCCI), Huashan Hospital (National Medical Center for Neurological Diseases) and Shanghai Mental Health Center (National Medical Center for Mental Diseases) on April 9.
&nbsp;
At the meeting, experts from the AI and clinical fields had active exchanges. Professor Xu Yifeng, Director of the Brain Health Institute at National Center for Mental Disorder, Professor Wang Zhen, Vice President of the Shanghai Mental Health Center, and Professor Li Weidong, Executive Director of the Institute of Psychology and Behavioral Science at Shanghai Jiao Tong University, attended the meeting.
&nbsp;
<h3>AI + wearable devices will help precision medicine solve the last-mile problem</h3>
&nbsp;
AI scientist Dr. Hu Pengwei pointed out that the application scenarios of AI in the medical field are extremely wide, and currently it mainly serves three purposes: reducing the burden of repetitive work, identifying traces and clues that are difficult to detect by human beings, and conducting clue analysis in complex environments. AI also has great potential in precision medicine. GPT&#8217;s summarizing and inductive ability has shown great strength in early examination and diagnosis, out-of-hospital emotional support and assistance, big data analytics and pattern identification. He predicts that with the combined innovation of AI technology and wearable smart devices, precision medicine is expected to solve the last-mile problem in the next three to five years.
&nbsp;
<h3>Deciphering invasive EEG to achieve infinite possibilities</h3>
&nbsp;
Professor Chen Liang, Deputy Director of Neurosurgery and Leader of Functional Neurosurgery at Fudan University-affiliated Huashan Hospital, focused on the construction of invasive EEG databases and the prospects of augmented AI technology in deciphering brain functions. Invasive EEG refers to the implantation of electrodes into the brain or placing them on the surface of the brain to obtain EEG data with high signal-to-noise ratio. This type of data is of critical value in the fields of neuroscience and neurosurgery, as neuronal firing is the most basic form of nerve cell activities. By implanting high-density surface or deep electrodes, researchers expect to collect more data on EEG activities. In the case of Parkinson&#8217;s disease patients, for example, clinicians urgently need to find the best treatment options through a large number of intracranial stimulation experiments, which is burdensome for patients. He hopes that augmented AI technology will complete the time-consuming and repetitive work previously done by human beings to address unanswered scientific questions, including epilepsy traceability warning and consciousness transformation.
&nbsp;
<h3>A prospective key role in treating AD</h3>
&nbsp;
Professor Yu Jintai, Deputy Director of the Department of Neurology at Fudan University-affiliated Huashan Hospital and Leader of cognitive disorders research at the National Medical Center for Neurological Diseases, pointed out that to achieve early diagnosis of AD, it is necessary to set up large cohorts, especially community-based cohorts, in order to identify patients in the preclinical stage. The GPT model holds great potential in disease management which will improve medical automation such as the construction of a disease management platform to achieve individualized patient condition assessment, automated analysis reporting, intelligent follow-up Q&amp;A, and other functions. He also mentioned that GPT faces many challenges in the field of AD diagnosis and research, such as the lack of high-quality medical data, data security issues, and the reliance of effective answers on the training data. However, he also believes that AI is expected to play a key role in the field of AD by continuously deepening research and practice.
&nbsp;
<h3>Contributive to dream decoding</h3>
&nbsp;
Professor Yu Huan, Executive Director of the Sleep Disorders Clinic at Fudan University-affiliated Huashan Hospital, emphasized the impact of sleep disorders on quality of life, such as cerebrovascular accidents and dementia. Currently, polysomnography (PSG) is the standard technique for sleep disorder diagnosis, but it is costly and inefficient. Therefore, researchers are looking forward to improving diagnostic methods through artificial intelligence technology. Professor Yu Huan introduced the application of dream research in the field of sleep disorders, such as improving memory by controlling dreams. There are more than 150 methods for encoding and calculating dreams, and the researchers hope to design a more practical research tool with the help of AI technology; at the same time, mobile apps are developed to encourage individuals to record and share their dreams, so as to carry out dream research more relevant to daily life.
&nbsp;
<h3>Robot-assisted consultation and diagnosis for depression patients</h3>
&nbsp;
According to Associate Professor Wu Mengyue of the Department of Computer Science and Engineering at Shanghai Jiao Tong University, the development of a robot based on human-computer conversations for clinical consultations on depression, as well as the use of speech and language features to build a knowledge graph of relevant symptoms and mental illnesses, is the future direction of early diagnosis and treatment of depression.
&nbsp;
According to her, the diagnosis of many mental illnesses relies heavily on face-to-face clinical interviews and conversations. Theoretically, AI models should be able to learn this skill as well. Meanwhile, speech and language are already used as objective biomarkers in the DSM-5 diagnostic manual to diagnose mental illnesses including depression. The development of a robot based on human-computer conversations for depression diagnosis would allow such conversations to be as accurate as the description of symptoms obtained by doctors through in-depth communications with patients. She also introduces how to treat speech and language features as computable and transferable methods, as well as the creation of knowledge graphs of relevant symptoms and diseases through the self-expressions of patients, which will provide new methodologies for detecting a wide range of diseases.
&nbsp;
<h3>New explorations of early diagnosis and treatment of depression</h3>
&nbsp;
Professor Peng Daihui, Director of the Department of Mood Disorders of Shanghai Mental Health Center, introduced that the major research project he is leading &#8212; “Prospective Clinical Cohort Study of Depression”, which aims to collect nationwide data from patients with depression, and to create a large-scale, multi-center, normalized and standardized long-term case database.
&nbsp;
Currently, the team has initially constructed a diagnostic and typing model of the brain function network for depression by combining brain imaging and clinical neuropsychological assessment. They propose to further apply digital phenotyping technology, including audio, video, EEG and eye movement, and other multidimensional stereoscopic big data for feature extraction, screening, and modeling. He pointed out that such multidimensional stereoscopic big data may improve the accuracy rate of depression diagnosis, optimize screening and assessment methods as well as the prediction of risk events; the combination of big data and artificial intelligence technologies has great potential in providing sensitive and specific diagnostic and treatment solutions for patients.
&nbsp;
<h3>Unlocking genetic secrets and probing knowledge graphs</h3>
&nbsp;
Professor Lin Guaning of the School of Biomedical Engineering at Shanghai Jiao Tong University demonstrated the results achieved by applying GPT technology in the fields of mental health and brain science research through continuous optimization of GPT training and rule setting.
&nbsp;
In the area of stress, depression and suicide risk detection, the research team has improved the accuracy of GPT through the prompt engineering and has been able to initially achieve accurate classification and prediction. The team also successfully extracted structured information from unstructured text, and by providing a prescribed schema for the GPT, achieved the standardized storage of such information in the database, which can be valuable for future research and clinical practice. Despite the challenges encountered in the application process, such as API interface limitations, Professor Lin believes that large language models like GPT will play an increasingly important role in mental health and brain science research, and will soon have the ability to process data other than textual language, such as multimodal data from imaging, EEG, and bio-omics, and to reason about the intrinsic logic amongst the data. This will revolutionize the existing research paradigm and drive the rapid development of the mental health and brain science research.

[:en]“In the face of the new AI technology that is developing by leaps and bounds, as a neurosurgeon, I'm not worried about losing my job, but more concerned about how to embrace AI to better serve patients, and to allow us doctors to get off work earlier. For example, AI-assisted clinical interrogations, brain image and EEG data analysis, and surgical scheme formulation; or examining medical issues from a completely new perspective of AI, overcoming the subjective shortcomings of studying human brain, and tackling brain diseases as soon as possible,” said Professor Mao Ying, President of Huashan Hospital and Director of the Translational Center of TCCI, in his opening speech at a seminar on AI-assisted solutions to brain diseases jointly hosted by the Tianqiao and Chrissy Chen Institute (TCCI), Huashan Hospital (National Medical Center for Neurological Diseases) and Shanghai Mental Health Center (National Medical Center for Mental Diseases) on April 9.

&nbsp;

At the meeting, experts from the AI and clinical fields had active exchanges. Professor Xu Yifeng, Director of the Brain Health Institute at National Center for Mental Disorder, Professor Wang Zhen, Vice President of the Shanghai Mental Health Center, and Professor Li Weidong, Executive Director of the Institute of Psychology and Behavioral Science at Shanghai Jiao Tong University, attended the meeting.

&nbsp;
<h3>AI + wearable devices will help precision medicine solve the last-mile problem</h3>
&nbsp;

AI scientist Dr. Hu Pengwei pointed out that the application scenarios of AI in the medical field are extremely wide, and currently it mainly serves three purposes: reducing the burden of repetitive work, identifying traces and clues that are difficult to detect by human beings, and conducting clue analysis in complex environments. AI also has great potential in precision medicine. GPT's summarizing and inductive ability has shown great strength in early examination and diagnosis, out-of-hospital emotional support and assistance, big data analytics and pattern identification. He predicts that with the combined innovation of AI technology and wearable smart devices, precision medicine is expected to solve the last-mile problem in the next three to five years.

&nbsp;
<h3>Deciphering invasive EEG to achieve infinite possibilities</h3>
&nbsp;

Professor Chen Liang, Deputy Director of Neurosurgery and Leader of Functional Neurosurgery at Fudan University-affiliated Huashan Hospital, focused on the construction of invasive EEG databases and the prospects of augmented AI technology in deciphering brain functions. Invasive EEG refers to the implantation of electrodes into the brain or placing them on the surface of the brain to obtain EEG data with high signal-to-noise ratio. This type of data is of critical value in the fields of neuroscience and neurosurgery, as neuronal firing is the most basic form of nerve cell activities. By implanting high-density surface or deep electrodes, researchers expect to collect more data on EEG activities. In the case of Parkinson's disease patients, for example, clinicians urgently need to find the best treatment options through a large number of intracranial stimulation experiments, which is burdensome for patients. He hopes that augmented AI technology will complete the time-consuming and repetitive work previously done by human beings to address unanswered scientific questions, including epilepsy traceability warning and consciousness transformation.

&nbsp;
<h3>A prospective key role in treating AD</h3>
&nbsp;

Professor Yu Jintai, Deputy Director of the Department of Neurology at Fudan University-affiliated Huashan Hospital and Leader of cognitive disorders research at the National Medical Center for Neurological Diseases, pointed out that to achieve early diagnosis of AD, it is necessary to set up large cohorts, especially community-based cohorts, in order to identify patients in the preclinical stage. The GPT model holds great potential in disease management which will improve medical automation such as the construction of a disease management platform to achieve individualized patient condition assessment, automated analysis reporting, intelligent follow-up Q&amp;A, and other functions. He also mentioned that GPT faces many challenges in the field of AD diagnosis and research, such as the lack of high-quality medical data, data security issues, and the reliance of effective answers on the training data. However, he also believes that AI is expected to play a key role in the field of AD by continuously deepening research and practice.

&nbsp;
<h3>Contributive to dream decoding</h3>
&nbsp;

Professor Yu Huan, Executive Director of the Sleep Disorders Clinic at Fudan University-affiliated Huashan Hospital, emphasized the impact of sleep disorders on quality of life, such as cerebrovascular accidents and dementia. Currently, polysomnography (PSG) is the standard technique for sleep disorder diagnosis, but it is costly and inefficient. Therefore, researchers are looking forward to improving diagnostic methods through artificial intelligence technology. Professor Yu Huan introduced the application of dream research in the field of sleep disorders, such as improving memory by controlling dreams. There are more than 150 methods for encoding and calculating dreams, and the researchers hope to design a more practical research tool with the help of AI technology; at the same time, mobile apps are developed to encourage individuals to record and share their dreams, so as to carry out dream research more relevant to daily life.

&nbsp;
<h3>Robot-assisted consultation and diagnosis for depression patients</h3>
&nbsp;

According to Associate Professor Wu Mengyue of the Department of Computer Science and Engineering at Shanghai Jiao Tong University, the development of a robot based on human-computer conversations for clinical consultations on depression, as well as the use of speech and language features to build a knowledge graph of relevant symptoms and mental illnesses, is the future direction of early diagnosis and treatment of depression.

&nbsp;

According to her, the diagnosis of many mental illnesses relies heavily on face-to-face clinical interviews and conversations. Theoretically, AI models should be able to learn this skill as well. Meanwhile, speech and language are already used as objective biomarkers in the DSM-5 diagnostic manual to diagnose mental illnesses including depression. The development of a robot based on human-computer conversations for depression diagnosis would allow such conversations to be as accurate as the description of symptoms obtained by doctors through in-depth communications with patients. She also introduces how to treat speech and language features as computable and transferable methods, as well as the creation of knowledge graphs of relevant symptoms and diseases through the self-expressions of patients, which will provide new methodologies for detecting a wide range of diseases.

&nbsp;
<h3>New explorations of early diagnosis and treatment of depression</h3>
&nbsp;

Professor Peng Daihui, Director of the Department of Mood Disorders of Shanghai Mental Health Center, introduced that the major research project he is leading -- “Prospective Clinical Cohort Study of Depression”, which aims to collect nationwide data from patients with depression, and to create a large-scale, multi-center, normalized and standardized long-term case database.

&nbsp;

Currently, the team has initially constructed a diagnostic and typing model of the brain function network for depression by combining brain imaging and clinical neuropsychological assessment. They propose to further apply digital phenotyping technology, including audio, video, EEG and eye movement, and other multidimensional stereoscopic big data for feature extraction, screening, and modeling. He pointed out that such multidimensional stereoscopic big data may improve the accuracy rate of depression diagnosis, optimize screening and assessment methods as well as the prediction of risk events; the combination of big data and artificial intelligence technologies has great potential in providing sensitive and specific diagnostic and treatment solutions for patients.

&nbsp;
<h3>Unlocking genetic secrets and probing knowledge graphs</h3>
&nbsp;

Professor Lin Guaning of the School of Biomedical Engineering at Shanghai Jiao Tong University demonstrated the results achieved by applying GPT technology in the fields of mental health and brain science research through continuous optimization of GPT training and rule setting.

&nbsp;

In the area of stress, depression and suicide risk detection, the research team has improved the accuracy of GPT through the prompt engineering and has been able to initially achieve accurate classification and prediction. The team also successfully extracted structured information from unstructured text, and by providing a prescribed schema for the GPT, achieved the standardized storage of such information in the database, which can be valuable for future research and clinical practice. Despite the challenges encountered in the application process, such as API interface limitations, Professor Lin believes that large language models like GPT will play an increasingly important role in mental health and brain science research, and will soon have the ability to process data other than textual language, such as multimodal data from imaging, EEG, and bio-omics, and to reason about the intrinsic logic amongst the data. This will revolutionize the existing research paradigm and drive the rapid development of the mental health and brain science research.[:zh]“面对发展一日千里的AI新技术，作为一名神经外科大夫，我并不担心会失业，更关心如何拥抱AI，更好地服务患者，还能让我们医生早点下班。比如AI辅助问诊、辅助分析大脑影像和脑电数据，制定手术方案等；再比如从AI全新的视角审视，打破人脑研究人脑的主观性障碍，早日攻克脑疾病。”4月9日，由天桥脑科学研究院（TCCI）携手华山医院（国家神经疾病医学中心）、上海市精神卫生中心（国家精神疾病医学中心）联合主办的AI助力攻克脑疾病研讨会上，华山医院院长、TCCI转化中心主任毛颖教授开门见山地说。

&nbsp;

会上，来自AI、临床领域的专家进行了积极交流。国家精神疾病医学中心脑健康研究院院长徐一峰教授、上海市精神卫生中心副院长王振教授、上海交通大学心理与行为科学研究院执行院长李卫东教授等参加了会议。

&nbsp;
<h3>AI+可穿戴设备，</h3>
<h3>精准医疗即将打通最后一公里</h3>
&nbsp;

人工智能科学家胡鹏伟博士指出，AI在医疗领域的应用场景极为广泛，目前主要实现三大功能：减轻重复劳动负担、识别人工难以察觉的痕迹和线索，以及在复杂环境中进行线索分析。AI在精准医疗方面也有巨大潜力，GPT的总结与归纳能力在早期检查与诊断、院外情感支持及辅助、大数据分析与模式识别等方面已经显现出强大实力。他预测，凭借AI技术和可穿戴智能设备等的结合创新，精准医疗有望在3～5年内完成它的最后一公里。

&nbsp;
<h3>解密侵入式脑电，塑造无限可能</h3>
&nbsp;

复旦大学附属华山医院神经外科副主任、功能神经外科带头人陈亮教授重点介绍了侵入性脑电数据库的建设以及增强AI技术在脑功能破译中的应用前景。侵入性脑电指的是将电极植入大脑或置于大脑表面，以获取高信噪比的脑电数据。这类数据在神经科学和神经外科领域具有至关重要的价值，因为神经元放电是神经细胞最基本的活动方式。通过建立高密度表面或深部电极，研究人员期望收集更多关于脑电活动的数据。以帕金森病患者为例，临床医生迫切需要通过大量颅内刺激实验来寻找最佳治疗方案，但这种方法对患者造成的负担较重。他希望借助增强AI技术完成耗时且重复性较高的工作，能协助解决尚未解答的科学问题，包括癫痫溯源预警和意识转化。

&nbsp;
<h3>有望在攻克AD中发挥关键作用</h3>
&nbsp;

复旦大学附属华山医院神经内科副主任、国家神经疾病医学中心认知障碍方向带头人郁金泰教授指出，为实现AD早期诊断，有必要建立大型队列，尤其是社区队列，以便识别临床前阶段的患者。GPT模型在疾病管理方面的潜力，如搭建疾病管理平台，实现患者个体化病情评估、自动化分析报告、智能随访问答等功能，以提高医疗自动化水平。他也提到，GPT在AD诊疗与研究领域所面临的诸多挑战，如高质量医疗数据的缺乏、数据安全性问题、回答实效性受训练数据影响等。但他同时认为，通过不断深化研究与实践，AI有望在AD领域发挥关键作用。

&nbsp;
<h3>期待助力解码梦境</h3>
&nbsp;

复旦大学附属华山医院睡眠障碍诊疗中心执行主任于欢教授强调了睡眠障碍对生活质量的影响，如脑血管意外、痴呆等。目前，多导睡眠监测是睡眠障碍诊断的标准技术，但其成本高且效率低。因此，研究者们期待通过人工智能技术改进诊断方法。于欢教授介绍了梦境研究在睡眠障碍领域的应用，如通过控制梦境提高记忆力。目前已有超过150种编码和计算梦境的方法，研究者们希望借助AI技术制作一个实用性更强的研究工具；同时，开发移动客户端以鼓励个体记录和分享自己的梦境，从而进行更贴近日常生活的梦境研究。

&nbsp;
<h3>打造抑郁症问诊人机对话助手</h3>
&nbsp;

上海交通大学计算机科学与工程系吴梦玥副教授认为，开发基于人机对话的抑郁症问诊机器人、以及利用语音和语言特征构建症状与精神疾病知识图谱，是未来抑郁症早诊早治的方向。

&nbsp;

她说，很多精神疾病的诊断主要依赖于面对面的问诊和交谈，理论上，模型也应该能够学会这个技能。同时，语音和语言作为客观生物标记物，在DSM-5诊断手册中已经被用于诊断抑郁症等精神疾病。开发基于人机对话的抑郁症问诊机器人，通过深度交流，人机对话能够得到与医生所得到的同样精确的症状描述。她还介绍了如何将语音和语言特征作为可计算、可迁移的方式，以及通过患者的自我表达建立症状和疾病的知识图谱，为多种疾病检测提供了新的思路。

&nbsp;
<h3>抑郁症早诊早治新探索</h3>
&nbsp;

上海市精神卫生中心心境障碍科主任彭代辉教授介绍，他正在领导的重大科研项目“抑郁症的前瞻性临床队列研究”旨在收集全国范围内的抑郁症患者数据，创建一个多中心、规范化、标准化的大规模长期病例数据库。

&nbsp;

目前，团队通过脑影像学和临床神经心理评估两个维度相结合初步构建了抑郁症脑功能网络诊断与分型模型。他们拟进一步运用数字表型技术，包括音频、视频、脑电和眼动等多维度立体大数据进行特征提取、筛选和建模。他指出，这种多维度立体大数据可能提高抑郁症诊断的精准度，优化筛查评估方法以及风险事件预测；大数据与人工智能技术相结合，在为患者提供敏感和特异的诊疗方案中有着巨大潜力。

&nbsp;
<h3>解锁基因秘密，挖掘知识图谱</h3>
&nbsp;

上海交通大学生物医学工程学院林关宁教授展示了通过持续优化GPT的训练和规则设定，将GPT技术应用于心理健康和脑科学研究领域所取得的成果。

&nbsp;

在压力、抑郁症和自杀风险检测方面，研究团队通过改进提示工程（Prompt Engineering），实现了GPT准确性的提升，已能初步实现准确的分类和预判。团队还成功地从非结构化文本中提取了结构化信息，通过为GPT提供规定模式，实现了将这些信息规范地存储在数据库中，为未来研究和临床实践提供了宝贵数据。尽管在应用过程中遇到了一些挑战，如API接口局限性等，但林关宁教授相信像GPT等大型语言模型在心理健康和脑科学研究领域将发挥越来越重要的作用，很快将有能力处理除文本语言之外的数据，如影像、脑电、生物组学等多模态的数据，并推理数据之间的内在逻辑。这将为现有的科研范式带来革命性变革，并推动心理健康和脑科学研究领域的快速发展。[:]

AI helps tackle brain diseases, reducing workload for doctors

[:en]AI helps tackle brain diseases, reducing workload for doctors[:zh]AI助力，早日攻克脑疾病，还能让医生早点下班[:]

Embodied Intelligence: The Pathway to the Future of Human Wisdom

[:en]Embodied Intelligence: The Pathway to the Future of Human Wisdom[:zh]具身智能：通向人类智慧的未来之路[:]

From 'Emergent Intelligence' to 'Superhuman': Envisioning the Pinnacle of AGI

[:en]From 'Emergent Intelligence' to 'Superhuman': Envisioning the Pinnacle of AGI[:zh]从“智能涌现”到“超人类”，通往AGI巅峰的终极设想[:]

[:en]
	&nbsp;


<h2>
	1. The Long and Arduous Road of Drug Development
</h2>


	Alzheimer&#39;s disease (AD), the most prevalent form of dementia, is rapidly evolving into a global public health crisis. Current statistics indicate that approximately 40 million people worldwide suffer from AD, with forecasts predicting an increase to 152 million by 2050. As patient numbers grow, the progressive decline in cognitive and memory functions significantly affects patients and their families, posing a substantial burden on socio-economic systems.



	&nbsp;



	In response to AD&rsquo;s devastation, the medical community has persistently pursued solutions. Over recent decades, significant progress has been made in understanding AD&rsquo;s pathogenesis, with various hypotheses targeting proteins such as beta-amyloid (A&beta;) and tau. Although some promising drugs have entered clinical trials, to date, very few have been approved for the market, and these primarily provide symptomatic relief rather than a cure.



	&nbsp;



	Among the popular monoclonal antibody drugs, Aduhelm was launched in 2022 but has since been gradually abandoned by its developers due to controversial effectiveness and safety concerns. Conversely, Leqembi, approved in 2023, has been shown to slow cognitive decline by 27% within 18 months in early-stage AD patients. However, its high cost and limited insurance coverage mean only about 398,000 patients worldwide may receive it by 2026.



	&nbsp;



	Unlike hypertension patients who can manage their condition with antihypertensive drugs, the hope of curing AD with a few pills remains unrealized. Yet, are there other options? Recently, a novel non-pharmacological treatment has emerged quietly: 40Hz audiovisual stimulation therapy. This therapy, featuring a unique mechanism and easy operation, has shown promise in animal studies and preliminary clinical trials, offering new hope in the fight against AD.



	&nbsp;


<h2>
	2. The 40Hz Audiovisual Stimulation: A New Pathway
</h2>


	Our brain is composed of approximately 86 billion neurons that work together like a vast symphony orchestra, performing complex cognitive functions. Neurons communicate through electrical signals, analogous to how orchestra members coordinate with the conductor and each other. During a performance, various musical instruments harmonize, creating a pleasing symphony. Similarly, in the brain, many neurons firing synchronously produce oscillations at specific frequencies, known as &quot;neural oscillations.&quot;



	<img alt="" class="aligncenter size-full wp-image-1901" height="414" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture1.png" width="784" />



	▷&nbsp;Different bands of neural oscillations. Figure source: Pandey, Pankaj, Richa Tripathi, and Krishna Prasad Miyapuram. &quot;Classifying oscillatory brain activity associated with Indian Rasa s using network metrics.&quot; Brain Informatics 9.1 (2022): 15.



	&nbsp;



	Neural oscillations, like musical notes, occur at different frequencies. Researchers classify these into several bands: &delta; (1-4Hz), &theta; (4-8Hz), &alpha; (8-12Hz), &beta; (12-30Hz), and &gamma; (30-100Hz). Each band is linked to specific brain functions. For example, &alpha; waves are present during wakeful relaxation, whereas &gamma; waves are associated with higher cognitive processes such as perception, memory, and attention. In Alzheimer&rsquo;s disease (AD) patients, &gamma; oscillations are often disrupted, akin to a disjointed climactic movement in a musical performance. Studies utilizing electroencephalography and magnetoencephalography have shown significantly reduced &gamma; oscillations in both AD patients and animal models, indicating a decline in neuronal synchrony. This disruption impacts information transmission and integration across different brain regions, playing a crucial role in the progression of AD.



	<img alt="" class="aligncenter size-full wp-image-1900" height="270" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture2.png" width="480" />



	▷&nbsp;Differences in &gamma;&nbsp;rhythms between healthy individuals and AD patients. Figure source: https://brainhealth.org.uk/?p=624



	&nbsp;



	So, can we restore the disrupted &gamma; oscillations in AD patients&rsquo; brains, allowing neurons to resume their rhythmic activity and play a harmonious symphony? The answer may be surprisingly simple: applying external 40Hz stimulation appears effective!



	&nbsp;



	In 2016, a groundbreaking study led by Professor Li-Huei Tsai from MIT, published in the prestigious journal &quot;Nature,&quot; revealed significant therapeutic effects of 40Hz light stimulation in treating AD in mice. This discovery was based on AD mouse models that, although not yet displaying evident amyloid plaques, showed clear disruptions in &gamma; rhythms closely related to higher cognitive functions like perception, learning, and memory.



	&nbsp;



	Researchers used optogenetic techniques to activate a key type of interneuron in the mouse brain. When these neurons were stimulated at a 40Hz frequency, the mice exhibited a significant reduction in the pathogenic amyloid-beta (A&beta;) protein, an effect not observed at other frequencies. Further analysis showed that 40Hz stimulation not only suppresses A&beta; production but also enhances phagocytic clearance by microglia. Stimulated microglia become more active, converging on plaques to perform a &quot;clean-up&quot; function, suggesting that regulating &gamma; rhythms could effectively combat AD through a dual mechanism.



	&nbsp;



	Although optogenetics is a compelling technology, it involves injecting viruses into the brain and implanting optical fibers, making it currently suitable only for animal experiments and not yet for clinical use. However, scientists have found that simply exposing AD mice to 40Hz flickering light for just one hour can reduce brain A&beta; levels by nearly 60%. After receiving light exposure for one hour daily over seven days, the amyloid plaques in their brains were also significantly reduced.



	&nbsp;



	What About Auditory Stimulation? In a 2019 study published in &quot;Cell,&quot; researchers explored the therapeutic effects of auditory stimulation further. The study found that 40Hz auditory stimulation reduced amyloid-beta (A&beta;) deposition in the auditory cortex of AD mice and synchronized neuronal activity in critical brain areas such as the hippocampus and medial prefrontal cortex. This synchronization with the 40Hz frequency exerted a therapeutic effect across a larger area. The mice also showed significant improvements in cognitive tasks, including spatial memory and object recognition.



	&nbsp;



	When sound and light were combined, the therapeutic effects were enhanced. This multisensory stimulation not only induced a more robust immune response but also significantly reduced A&beta; in the medial prefrontal cortex. Whole-brain three-dimensional imaging revealed that combined stimulation significantly reduced plaques throughout the entire cortical brain and decreased levels of hyperphosphorylated tau protein. These findings suggest that 40Hz audiovisual stimulation therapy affects not just the primary sensory areas but also crucial hubs involved in learning and memory such as the hippocampus and prefrontal cortex, improving cognitive functions in AD mice behaviorally.



	<img alt="" class="aligncenter size-full wp-image-1899" height="498" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture3.png" width="498" />



	▷&nbsp;Audiovisual stimulation treatment for AD mice. Figure source: Martorell, Anthony J., et al. &quot;Multi-sensory &gamma;&nbsp;stimulation ameliorates Alzheimer&rsquo;s-associated pathology and improves cognition.&quot; Cell 177.2 (2019): 256-271.



	&nbsp;



	Another study published in &quot;Neuron&quot; the same year investigated the mechanisms behind 40Hz light stimulation. Researchers discovered that 40Hz light stimulation enhanced &gamma; rhythms in other brain regions, including the hippocampus, visual cortex, somatosensory cortex, and prefrontal cortex. It also slowed the loss of neurons and synapses in these areas, improving the mice&#39;s spatial learning and memory capabilities.



	&nbsp;



	Molecular-level studies revealed several potential protective mechanisms: suppression of pro-inflammatory gene expression in microglia, reduction of neuroinflammation, upregulation of genes related to synaptic function in neurons, enhancement of synaptic plasticity, induction of neuroprotective factors, and reduction of neuronal DNA damage, thereby helping to maintain genomic stability.



	&nbsp;



	Notably, 40Hz stimulation regulates not only the pathological changes associated with Alzheimer&#39;s disease but also shows a similar regulatory effect in normally aging mice, suggesting that it could broadly improve neural functions.



	<img alt="" class="aligncenter size-full wp-image-1898" height="663" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture4.png" style="" title="" width="663" />



	▷&nbsp;Treatment mechanisms. Figure source: Adaikkan, Chinnakkaruppan, et al. &quot;&gamma;&nbsp;entrainment binds higher-order brain regions and offers neuroprotection.&quot; Neuron 102.5 (2019): 929-943.



	&nbsp;


<h2>
	3. Skepticism from Other Teams
</h2>


	As the academic community marveled at the remarkable efficacy of such a simple therapy, a study published in March 2023 by Gy&ouml;rgy Buzs&aacute;ki&#39;s team at New York University in Nature Neuroscience delivered a significant blow to the 40Hz treatment credibility.



	&nbsp;



	Their research involved recording neural activity in various brain regions, including the visual cortex, olfactory cortex, and hippocampus. The study found that 40Hz light stimulation did not induce &gamma; oscillations and had a minimal impact on deeper brain regions. Further experiments indicated that AD mice seemed to dislike and actively avoid the 40Hz flickering light.



	&nbsp;



	The researchers also examined the effect of 40Hz light stimulation on AD pathology. Despite employing multiple experimental methods, they observed no changes in the quantity of amyloid plaques, no significant differences in microglial morphology, and no statistical differences in the levels of A&beta;40 and A&beta;42 between the treatment and control groups. Although some pathological indicators showed improvement, these changes were not stable or consistent.



	&nbsp;



	Does this mean that 40Hz audiovisual stimulation therapy for AD is ineffective? Not necessarily. Numerous factors can influence the outcomes of biological experiments. Even if all data are accurate and reliable, differences in key experimental steps, mouse strains, or slight changes in the environment can lead to vastly different results. Additionally, the study focused only on molecular pathological changes within the brain and did not explore the impact of the 40Hz therapy on cognitive functions, which are clearly more crucial. As the authors themselves stated, their results should not be viewed as evidence against the therapy&#39;s effectiveness in treating AD, and further exploration is needed to develop the 40Hz audiovisual stimulation therapy effectively.



	&nbsp;


<h2>
	4. New Research Tops &quot;Nature&quot; Again
</h2>


	In response to ongoing skepticism, Professor Tsai&#39;s team published a groundbreaking study in &quot;Nature&quot; in March 2024, uncovering a novel therapeutic mechanism for the 40Hz stimulation therapy.



	&nbsp;



	For normal brain function, an efficient &quot;garbage disposal system&quot; is essential to continuously remove waste. The glymphatic system, a crucial component of this system, depends on the unobstructed flow of cerebrospinal fluid (CSF) and interstitial fluid. It has been found that abnormal &gamma; oscillations may disrupt this process, potentially exacerbating Alzheimer&#39;s disease (AD).



	&nbsp;



	The study discovered that just one hour of 40Hz audiovisual stimulation significantly increases the influx of CSF and the efflux of interstitial fluid in the cortical brain of AD mice. This effect is mediated by the key protein AQP4 on astrocytes. Inhibiting AQP4 with the small molecule TGN020 diminished the inflow of CSF and impaired the clearance of amyloid-beta (A&beta;). These findings were further validated in mice genetically modified to lack the AQP4 gene. Additional research demonstrated that 40Hz stimulation enhances CSF flow by promoting the polarization of AQP4 on the cell membrane.



	&nbsp;



	Further investigations have shown that arterial pulsation regulates CSF flow and that lymphatic vessels facilitate the drainage of CSF from the brain, also aiding in the transport of A&beta;. Experiments indicate that 40Hz audiovisual stimulation also boosts arterial pulsation and increases the diameter of lymphatic vessels in the brains of mice, improving A&beta; clearance.



	&nbsp;



	So, who coordinates all these processes? Subsequent experiments have revealed that these are regulated by VIPergic interneurons, which secrete the vasoactive intestinal peptide (VIP). Prior studies have indicated that the VIP pathway plays a role in regulating arterial diameter, astrocytic metabolism, and aquaporin transport, among other physiological processes. Inhibiting VIP neurons prior to 40Hz stimulation resulted in decreased A&beta; clearance, AQP4 polarization, and arterial pulsation.



	&nbsp;



	This research systematically elucidates a novel mechanism: 40Hz stimulation, via &gamma; oscillations, promotes the synergistic action of a neural-vascular-glia circuit centered around VIPergic interneurons, thereby accelerating AQP4-mediated CSF flow and the removal of pathogenic proteins. This multi-level regulatory process&mdash;from gene expression to cellular function and tissue level&mdash;reveals intrinsic connections between neural activity, CSF circulation, and AD pathology, further affirming the efficacy of 40Hz audiovisual stimulation in combating AD.



	&nbsp;


<h2>
	5. Rapid Clinical Translation
</h2>


	Building on these groundbreaking discoveries, Professor Li-Huei Tsai and her collaborators founded Cognito Therapeutics to officially begin the journey of transitioning 40Hz audiovisual therapy from the laboratory to clinical settings.



	&nbsp;



	At the 2021 Alzheimer&#39;s/Parkinson&#39;s Disease Congress, Cognito unveiled the results of a multicenter, randomized, controlled Phase II clinical trial (NCT-03556280). After six months of treatment, patients with mild to moderate Alzheimer&rsquo;s disease showed good tolerance to the company&#39;s 40Hz audiovisual stimulation device, named Gamma Sense. Patients in the treatment group, compared to those in the placebo or sham group, experienced an 84% reduction in cognitive decline, an 83% slowdown in the deterioration of daily living skills, and a 61% reduction in related brain atrophy.



	<img alt="" class="aligncenter size-large wp-image-1897" height="475" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture5-1024x949.png" style="" title="" width="512" />



	▷&nbsp;Clinical trial, showing a slowdown in cognitive decline.



	&nbsp;



	In the same year, the FDA designated Gamma Sense as a Breakthrough Device. Additional findings were published in the March 2024 issue of &quot;Frontiers in Neuroscience.&quot;



	&nbsp;



	Following the initial success of the Phase II trial, a Phase III clinical trial has been aggressively pursued. This trial aims to enroll over 500 patients who will receive daily audiovisual therapy for 12 consecutive months, with the trial&#39;s completion expected in 2025.



	<img alt="" class="aligncenter size-large wp-image-1896" height="341" src="https://admin.next-question.com/wp-content/uploads/2024/05/Picture6-1024x682.png" style="" title="" width="512" />



	▷&nbsp;40Hz audiovisual stimulation device, The Spectris. Source: COGNITO THERAPEUTICS.



	&nbsp;


<h2>
	6. Conclusion
</h2>


	Undoubtedly, 40Hz audiovisual stimulation therapy offers unique advantages. Unlike traditional pharmacological treatments, it modulates neural oscillations&mdash;a more upstream component&mdash;in a synergistic manner across multiple levels, providing neuroprotection and potentially slowing down the progression of Alzheimer&rsquo;s disease at various stages. As a non-invasive physical therapy, it is easy to administer, safe, and can be performed by patients at home. This convenience greatly improves patient compliance and reduces treatment costs.



	&nbsp;



	However, it is crucial to acknowledge that this therapy is still in the clinical trial phase. Early research has been limited by small sample sizes and short follow-up periods. Should Phase III clinical trials continue to demonstrate efficacy and long-term safety, and if personalized treatment adaptations are incorporated, 40Hz audiovisual stimulation therapy could emerge as a truly beneficial, non-invasive, safe, and affordable method to delay the progression of AD.



	&nbsp;



	It is also worth noting that 40Hz audiovisual stimulation therapy does not conflict with existing drug therapies; instead, it may enhance them. We can envision a scenario where patients receive regular injections of antibody medications while also using a daily head-mounted audiovisual stimulation device. This approach could enhance the effects of both treatments in a coordinated, precision medicine approach.



	&nbsp;



	While the prospects are promising, the path toward clinical translation is long and fraught with challenges. We look forward to ongoing efforts in both basic and clinical research that will further elucidate the mechanisms of the treatment, assess tolerance for long-term use, confirm the durability of the effects, and evaluate suitability across different populations. Through these endeavors, this innovative therapy may flourish and help humanity combat the formidable challenge of Alzheimer&rsquo;s disease.



	&nbsp;



	&nbsp;



	&nbsp;

[:zh]

[:]

Why is 40Hz Remarkably Effective?

[:en]Why is 40Hz Remarkably Effective?[:]

TCCI Hosts Second “AI for Brain Science” Conference focused on Data Generation for Medical AI Models
&nbsp;
On May 28, the Tianqiao and Chrissy Chen Institute (TCCI) hosted the second session of the “AI for Brain Science” series themed &#8212; “Data Generation Methods for AI Models and Their Implications for the Medical Field”. Chaired by Mengyue Wu, Associate Professor of the Department of Computer Science and Engineering of Shanghai Jiao Tong University, three young scientists shared their practices and views on breaking the bottleneck of data for large language models (LLM).
<h3></h3>
<h3>Self-training and self-distillation: developing proprietary GPT models efficiently</h3>
&nbsp;
An international study has determined that ChatGPT can answer cancer-related questions at a level already on par with the official answers provided by the US National Cancer Institute. ChatGPT however can only be accessed through restricted APIs and when it comes to personal healthcare, the public at large do not want to share private information with third-party companies.
&nbsp;
To resolve such difficulties, Canwen Xu, a PhD student at UC San Diego and collaborators from Sun Yat-sen University proposed a process that automatically generates a high-quality multi-round chat corpus by using ChatGPT to chat with itself. The conversational data can then be used to fine-tune and enhance the open-source large language model called LLaMA. As a result, they obtained a high-quality proprietary model called “Bai Ze” whose version 2.0 was launched just a few days ago. The name was inspired by a mythical beast in ancient China that can speak and understand the feelings of all things.
&nbsp;
According to Xu, Bai Ze did not learn anything new in the process, but simply extracted specific data from LLaMA and retained the powerful linguistic capabilities of ChatGPT to answer questions by bullet points or refuse to answer certain inappropriate questions. This has been professionally referred to as ‘distillation’. The concept of feedback self-distillation was also introduced, whereby ChatGPT is used as a coach to score and rank the answers provided by Bai Ze to further improve its performance.
&nbsp;
Xu thinks that Bai Ze is both economical and pragmatic as it acquired ChatGPT&#8217;s capabilities in a certain domain through automated knowledge distillation at a much lower cost. In the medical field, localized or private AI models will help ease privacy-related concerns and assist in diagnosis and treatment. In the future, perhaps everyone will have a personal AI assistant.
&nbsp;
<h3>A new data generation strategy: optimizing medical text mining</h3>
&nbsp;
Ruixiang Tang, a PhD student at Rice University and his collaborators have also proposed a new data generation strategy based on large language models. It has demonstrated better performance on classic medical text mining tasks such as named entity recognition (NER) and relationship extraction (RE).
&nbsp;
Since ChatGPT is capable of creative writing, it performs well in knowledge-intensive areas with little annotated data including healthcare, finance, and law. However, when it comes to medical text mining, they found that applying ChatGPT directly in the downstream tasks of processing medical text did not always achieve ideal results and might even cause privacy-related problems.
&nbsp;
To address this challenge, Tang and his team proposed a new strategy: using large language models to generate large amounts of medical data, and then training these data with smaller models. The experimental results show that this new strategy achieved better results than the direct application of large language models in downstream tasks. It has also significantly reduced potential privacy risks since the data is stored locally.
&nbsp;
They further pointed out that as open-source large language models are improving in terms of both quantity and quality, the gap between the text data generated by AI and those by human will be increasingly narrowed. It’ll be technically challenging to tell the difference. The two existing testing means are likely to be ineffective, whether it is black-box testing, which directly compares text data generated by large language models with those generated by humans (e.g. comparing the distribution of high-frequency words), or white-box testing, where developers will label the generated text. The ability to effectively detect whether the data is generated by GPT will influence users’ trust in LLM-based AI.
&nbsp;
<h3>What’s so unique about data generation in the large language model era?</h3>
&nbsp;
In history, how did scientists address the challenge of data scarcity without GPT? And what new trends have large language models set?
&nbsp;
Ruisheng Cao, a PhD student at Shanghai Jiao Tong University, reviewed research in automated data generation or augmentation based on deep learning models upon the dawning of LLM era. Deep learning is essentially a process of finding the mapping from an input x to an output y. Therefore, a large number of (x,y) data pairs are needed for training. In areas like healthcare where large amounts of authentic data are difficult to obtain, more (x,y) data pairs need to be generated manually.
&nbsp;
Cao breaks down data generation into three main modules. The first is for the generation of label (y). It aims at coupling the generated labels with the distribution of the authentic data for comparison. The second is responsible for the methods and constraints in generating the initial data (x). The third will ensure data quality once a complete data pair has been formed.
&nbsp;
Continuously growing in scale and capability, large language models are generating data with improving quality. Models trained by generated data can not only solve simple tasks, such as text classification, but also handle more complex tasks such as Q&amp;A sessions.
&nbsp;
Looking forward, Cao summarized several new trends in data generation in the era of large language models. The first is to build more generalized models to ensure their application in diverse tasks. This means that models need to be more adaptable and generalized. The second is to further refine the process from a task-specific perspective. For example, in the medical field, it is even possible to customize tasks for specific types of depression, thus offering more precise and personalized medical solutions. Last but not the least, the process of data generation and model training will be more integrated, and mandatory filtering will be gradually replaced by flexible control to ensure data quality.
&nbsp;
Research into data generation and its application are tapping into the potentials of AI models in various domains, especially the medical field.
&nbsp;
<div id="attachment_1089" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1089" class="wp-image-1089" src="https://admin.next-question.com/wp-content/uploads/2023/11/19-1-300x300.png" alt="" width="600" height="600" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/19-1-300x300.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/19-1-150x150.png 150w, https://admin.next-question.com/wp-content/uploads/2023/11/19-1.png 400w" sizes="(max-width: 600px) 100vw, 600px" />Scan the QR code to watch the replay</div>

[:en]TCCI Hosts Second “AI for Brain Science” Conference focused on Data Generation for Medical AI Models

&nbsp;

On May 28, the Tianqiao and Chrissy Chen Institute (TCCI) hosted the second session of the “AI for Brain Science” series themed -- “Data Generation Methods for AI Models and Their Implications for the Medical Field”. Chaired by Mengyue Wu, Associate Professor of the Department of Computer Science and Engineering of Shanghai Jiao Tong University, three young scientists shared their practices and views on breaking the bottleneck of data for large language models (LLM).
<h3></h3>
<h3>Self-training and self-distillation: developing proprietary GPT models efficiently</h3>
&nbsp;

An international study has determined that ChatGPT can answer cancer-related questions at a level already on par with the official answers provided by the US National Cancer Institute. ChatGPT however can only be accessed through restricted APIs and when it comes to personal healthcare, the public at large do not want to share private information with third-party companies.

&nbsp;

To resolve such difficulties, Canwen Xu, a PhD student at UC San Diego and collaborators from Sun Yat-sen University proposed a process that automatically generates a high-quality multi-round chat corpus by using ChatGPT to chat with itself. The conversational data can then be used to fine-tune and enhance the open-source large language model called LLaMA. As a result, they obtained a high-quality proprietary model called “Bai Ze” whose version 2.0 was launched just a few days ago. The name was inspired by a mythical beast in ancient China that can speak and understand the feelings of all things.

&nbsp;

According to Xu, Bai Ze did not learn anything new in the process, but simply extracted specific data from LLaMA and retained the powerful linguistic capabilities of ChatGPT to answer questions by bullet points or refuse to answer certain inappropriate questions. This has been professionally referred to as ‘distillation’. The concept of feedback self-distillation was also introduced, whereby ChatGPT is used as a coach to score and rank the answers provided by Bai Ze to further improve its performance.

&nbsp;

Xu thinks that Bai Ze is both economical and pragmatic as it acquired ChatGPT's capabilities in a certain domain through automated knowledge distillation at a much lower cost. In the medical field, localized or private AI models will help ease privacy-related concerns and assist in diagnosis and treatment. In the future, perhaps everyone will have a personal AI assistant.

&nbsp;
<h3>A new data generation strategy: optimizing medical text mining</h3>
&nbsp;

Ruixiang Tang, a PhD student at Rice University and his collaborators have also proposed a new data generation strategy based on large language models. It has demonstrated better performance on classic medical text mining tasks such as named entity recognition (NER) and relationship extraction (RE).

&nbsp;

Since ChatGPT is capable of creative writing, it performs well in knowledge-intensive areas with little annotated data including healthcare, finance, and law. However, when it comes to medical text mining, they found that applying ChatGPT directly in the downstream tasks of processing medical text did not always achieve ideal results and might even cause privacy-related problems.

&nbsp;

To address this challenge, Tang and his team proposed a new strategy: using large language models to generate large amounts of medical data, and then training these data with smaller models. The experimental results show that this new strategy achieved better results than the direct application of large language models in downstream tasks. It has also significantly reduced potential privacy risks since the data is stored locally.

&nbsp;

They further pointed out that as open-source large language models are improving in terms of both quantity and quality, the gap between the text data generated by AI and those by human will be increasingly narrowed. It’ll be technically challenging to tell the difference. The two existing testing means are likely to be ineffective, whether it is black-box testing, which directly compares text data generated by large language models with those generated by humans (e.g. comparing the distribution of high-frequency words), or white-box testing, where developers will label the generated text. The ability to effectively detect whether the data is generated by GPT will influence users’ trust in LLM-based AI.

&nbsp;
<h3>What’s so unique about data generation in the large language model era?</h3>
&nbsp;

In history, how did scientists address the challenge of data scarcity without GPT? And what new trends have large language models set?

&nbsp;

Ruisheng Cao, a PhD student at Shanghai Jiao Tong University, reviewed research in automated data generation or augmentation based on deep learning models upon the dawning of LLM era. Deep learning is essentially a process of finding the mapping from an input x to an output y. Therefore, a large number of (x,y) data pairs are needed for training. In areas like healthcare where large amounts of authentic data are difficult to obtain, more (x,y) data pairs need to be generated manually.

&nbsp;

Cao breaks down data generation into three main modules. The first is for the generation of label (y). It aims at coupling the generated labels with the distribution of the authentic data for comparison. The second is responsible for the methods and constraints in generating the initial data (x). The third will ensure data quality once a complete data pair has been formed.

&nbsp;

Continuously growing in scale and capability, large language models are generating data with improving quality. Models trained by generated data can not only solve simple tasks, such as text classification, but also handle more complex tasks such as Q&amp;A sessions.

&nbsp;

Looking forward, Cao summarized several new trends in data generation in the era of large language models. The first is to build more generalized models to ensure their application in diverse tasks. This means that models need to be more adaptable and generalized. The second is to further refine the process from a task-specific perspective. For example, in the medical field, it is even possible to customize tasks for specific types of depression, thus offering more precise and personalized medical solutions. Last but not the least, the process of data generation and model training will be more integrated, and mandatory filtering will be gradually replaced by flexible control to ensure data quality.

&nbsp;

Research into data generation and its application are tapping into the potentials of AI models in various domains, especially the medical field.

&nbsp;

[caption id="attachment_1089" align="aligncenter" width="600"]<img class="wp-image-1089" src="https://admin.next-question.com/wp-content/uploads/2023/11/19-1-300x300.png" alt="" width="600" height="600" /> Scan the QR code to watch the replay[/caption]

 [:zh]都说医疗、金融等专业领域的语料数据稀缺，制约大模型AI的发展，那能不能让两个ChatGPT对聊，聊出点数据出来？

&nbsp;

5月28日，天桥脑科学研究院（Tianqiao &amp;Chrissy Chen Institute, TCCI）主办AI For Brain Science系列会议第二期——“面向AI模型的数据生成方法及其对医疗领域的启示”。在上海交通大学计算机科学与工程系副教授吴梦玥主持下，三名青年科学家分享了关于破解大规模语言模型（LLM）数据瓶颈的看法和实践。

&nbsp;
<h3>自对话和自蒸馏训练——快速构建专属GPT</h3>
&nbsp;

国际上一项研究评估指出，ChatGPT回答癌症相关问题的水平已经与美国国家癌症研究所的官方回答持平。然而，ChatGPT只能通过受限的API进行访问。涉及到个人医疗，人们也普遍不希望将自己的隐私信息分享给第三方公司。

&nbsp;

针对这样的难题，加州大学圣迭戈分校博士生许灿文和中山大学团队的合作者提出了一种能自动生成高质量多轮聊天语料库的流程，利用ChatGPT与其自身进行对话，生成对话数据，再基于产生的对话数据调优、增强开源的大型语言模型LLaMA。他们从而获得了高质量的专属模型“白泽”，并在数天前推出了2.0版本。这个名字的灵感来源是中国古代传说中的一种神兽，“能言语，达知万物之情”。

&nbsp;

许灿文介绍道，白泽在这个过程中并没有学会新的知识，只是提取了大模型中的特定数据，并且保留了ChatGPT分点作答、拒绝回答等强大的语言能力。这在专业上被比喻为一种“蒸馏”。进一步地，他们提出了反馈自蒸馏的概念，即利用ChatGPT当教官，对白泽回答的结果进行评分排序，从而进一步提高了白泽模型的性能。

&nbsp;

许灿文认为，白泽通过自动化的知识蒸馏，在特定领域达到ChatGPT的能力，成本却远远低于ChatGPT，兼具经济意义和实用意义。在医疗领域，本地化或私有化建构的模型将有利于消除隐私顾虑，辅助患者诊疗。未来也许每个人都将有自己的专属AI助手。

&nbsp;
<h3>一种新的数据生成策略：大模型优化医疗文本挖掘</h3>
&nbsp;

莱斯大学博士生唐瑞祥和合作者同样基于大模型提出了一种新的数据生成策略，并在命名实体识别（NER）、关系提取（RE）等经典的医疗文本挖掘任务上取得了更好的表现。

&nbsp;

ChatGPT具有创造性的写作能力，在医疗、金融、法律等标注数据很少的领域以及知识密集型领域表现出色。然而，具体到医疗文本挖掘，他们发现将ChatGPT直接应用大型模型处理医疗文本的下游任务，表现并不总是优秀，也可能引发隐私问题。

&nbsp;

唐瑞祥等提出了一种新策略：利用大型模型生成大量医疗数据，再通过小型模型对这些数据进行训练。实验结果显示，相较于直接利用大型模型执行下游任务，这一新策略能够取得更出色的效果，同时因为模型数据在本地，也大幅降低了潜在的隐私风险。

&nbsp;

他们进一步指出，随着开源大模型数量的增加和大模型能力的提升，其产生的文本数据与人类产生的文本数据的差别将越来越小，发展检测二者差别的技术手段将是一项富有挑战性的工作。现有的两种检测手段，无论是黑盒检测——直接比较大模型生成的文本数据与人类生成的文本数据（比如比较高频词分布），还是白盒检测——开发者在生成文本上做标签，在未来都可能失效。能否有效地检测出数据是不是GPT生成的，将影响到广大用户对大模型AI的信任程度。

&nbsp;
<h3>大模型时代的数据生成有什么不一样？</h3>
&nbsp;

那么，从历史演变的角度来看，在没有GPT的时代，科学家们如何解决数据稀缺难题？大模型又带来了哪些新趋势？

&nbsp;

上海交通大学博士生曹瑞升对大模型时代来临前夕，基于深度学习模型进行自动化数据生成或增广方面的研究，尤其是反向生成进行了回顾性的总结。深度学习本质上是一种找出从输入x到输出y的映射的过程，所以需要大量的(x,y)数据对来训练。在医疗这样不容易获得大量真实数据的领域，就需要人为生成更多的(x,y)数据对。

&nbsp;

曹瑞升将数据生成拆解为三个主要模块。第一个是针对标签（y）的生成，介绍如何对将生成的标签与真实数据的分布进行耦合比较。第二个模块是在生成数据时，介绍生成初始数据（x）的方法和限制。第三个模块是在形成完整的数据(x,y)对之后，应该如何保证数据质量。

&nbsp;

随着大语言模型规模的不断增大和能力的不断提升，其生成的数据质量也越来越高。这种生成数据所训练得到的模型不仅可以解决简单的任务，如文本分类，还可以应对问答等更加复杂的任务。

&nbsp;

展望未来，曹瑞升总结了数据生成在大模型时代的几大新趋势。首先是构建更加通用的模型，以确保其能够应用于多样化的任务。这意味着模型需要具备广泛的适应性和泛化能力。其次是从特定任务出发，进一步精细化地处理。例如，在医疗领域，甚至可以针对特定类型的抑郁症进行专业化的任务处理，提供更加精准和个性化的解决方案。最后，数据生成和模型训练的过程将从分离走向融合，而为了保证数据质量的硬性过滤也将逐渐被软性控制所取代。

&nbsp;

数据生成研究与应用的发展，为大模型AI走向各个专业领域，尤其是医疗领域提供广阔的可能性。

&nbsp;
<h3>结语</h3>
&nbsp;

TCCI致力于支持全球范围内的脑科学交流，仅2022年就主办、合办、支持了近200场会议，遍及北美、亚洲、欧洲和大洋洲。AI For Brain Science系列会议致力于促进AI与脑科学研究人员的讨论合作，将持续聚焦领域内的数据瓶颈和关键痛点，为大模型AI的未来突破提供创新土壤，促进前沿AI技术在脑科学领域发挥更大的价值。

&nbsp;

TCCI由盛大集团创始人，中国网络游戏、网络文学行业开创者陈天桥、雒芊芊夫妇出资10亿美元创建，聚焦AI＋脑科学，支持、推进全球范围内脑科学研究，造福人类。TCCI一期投入5亿元人民币支持中国脑科学研究，与上海周良辅医学发展基金会合作成立上海陈天桥脑健康研究所，与华山医院、上海市精神卫生中心等建立战略合作，设立了应用神经技术前沿实验室、人工智能与精神健康前沿实验室。在国际上，TCCI与加州理工学院合作成立TCCI加州理工研究院。

&nbsp;[:]

Using ChatGPT to Generate Data for Medical AI Models

[:en]Using ChatGPT to Generate Data for Medical AI Models[:zh]医疗AI缺数据，何不让ChatGPT自己聊出来？天桥脑科学研究院举办青年科学家研讨会[:]

How many of these 20 groundbreaking neuroscience researches of the year are you familiar with?

[:en]How many of these 20 groundbreaking neuroscience researches of the year are you familiar with?[:]

Commemorating Wittgenstein: How Does He Influence the Future of Large Language Models?

[:en]Commemorating Wittgenstein: How Does He Influence the Future of Large Language Models?[:]

The "Thousand Brains Project" Officially Launched, Entering the AI Arena

[:en]The "Thousand Brains Project" Officially Launched, Entering the AI Arena[:]

[:en]
	&nbsp;



	&nbsp;



	The continuity of memory is like a forward-flowing river of time, carrying the journey of eternal moments.



	Spanish surrealist Salvador Dal&iacute; illustrated this concept in his famous work &quot;The Persistence of Memory&quot;: soft clocks hanging eternally from the branches of life. However, in the era of quantum mechanics, this ever-contemplative artist deconstructed his creation in a work titled &quot;The Chromosome of a Highly Coloured Fish&#39;s Eye Starting the Harmonious Disintegration of the Persistence of Memory,&quot; later known as &quot;The Disintegration of the Persistence of Memory.&quot;



	<img alt="" class="aligncenter size-full wp-image-1981" height="566" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture13.png" width="745" /> <img alt="" class="aligncenter size-full wp-image-1982" height="839" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture9.png" width="1000" />



	▷&nbsp;Top: &quot;The Persistence of Memory&quot;; Bottom: Initially named &quot;The Chromosome of a Highly Coloured Fish&#39;s Eye Starting the Harmonious Disintegration of the Persistence of Memory,&quot; later known as &quot;The Disintegration of the Persistence of Memory.&quot; Source: Wikipedia.



	&nbsp;



	Yet, Dal&iacute; might not have anticipated that the fragmentation of DNA contained in chromosomes could be a significant source of new memories. Recently, Vladimir Jovasevic and his colleagues at Northwestern University&#39;s Feinberg School of Medicine provided an important explanation in an article in Nature: the damage and repair of double-stranded DNA in certain discrete neurons can initiate innate immunity, becoming a source of memory.



	<img alt="" class="aligncenter size-large wp-image-1980" height="479" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture14-1024x479.png" width="1024" />



	▷&nbsp;Vladimir Jovasevic, Vladimir, et al. &quot;Formation of memory assemblies through the DNA-sensing TLR9 pathway.&quot; Nature 628.8006 (2024): 145-153.



	&nbsp;



	1. Starting from Semon: Memory Exists in Specific Clusters of Cells



	Does the brain store memories? This seemingly obvious question in contemporary neuroscience was hotly debated among leading scholars in the field of learning and memory a century ago. For some, it was clear that memory was physically represented in the brain, while others believed it was stored in a more spiritualized mind.



	In 1904, Richard Semon wrote two monumental works (Die Mneme and Die Mnemischen Empfindungen) proposing a physical theory of human memory.



	He introduced the term &quot;engram,&quot; suggesting that the brain&#39;s state and structure respond physically to a person&#39;s experiences. This response implies the existence of a &quot;neural substrate&quot; in the brain for storing and reproducing memories. Once formed, the engram enters a dormant state and can be reactivated when past events recur.



	The development of experimental science gradually confirmed the physical existence of this concept.



	Karl Spencer Lashley, a pioneer of physiological psychology, initiated the search for the &quot;engram.&quot; By lesioning different parts of the cerebral cortex and observing the behavior of rats in mazes, he ultimately failed to find the engram. Yet, despite his frustration, Lashley remained convinced of the engram&#39;s existence&mdash;he believed it was distributed: &quot;It&#39;s not about finding where the trace is, but rather where it isn&#39;t.&quot;[1]



	A significant leap in engram research occurred when Lashley&#39;s student, psychologist, and memory theorist Donald O. Hebb proposed the cell assembly theory, similar to Semon&#39;s engram complex. Hebb hypothesized that a cell assembly forms among interconnected cells that are simultaneously active during stimulation. Sufficient activity within the cell assembly induces growth and metabolic changes, strengthening the connections between these cells. These synaptic and metabolic changes (potentially including alterations in intrinsic neuronal excitability) affect the function of the cell assembly.



	With the advent of more advanced technologies, the search for memory cells underwent a profound revival. By using labeling tracers and targeted interventions on individual neurons (activating or ablating them), researchers finally identified a group of neurons: these neurons exhibit a rapid high expression of immediate early genes during memory formation. Ablating these neurons prevents memory formation (necessity), while artificially activating them induces memory formation (sufficiency)&nbsp;[2].



	The discovery of engram cells demonstrates that memory indeed resides in specific populations of neurons and their interactive circuits.



	Now, neuroscientists are working to determine how neuronal groups in the hippocampus respond to stimuli that induce memory.



	&nbsp;



	2. DNA Damage and Repair in Engram Cells



	&quot;Engrams&quot; refer to the enduring offline physical and/or chemical changes induced by learning, which form the basis of newly established memory associations. Neurons with these physical memory traces (memory engram cells) are typically those that express immediate early genes during specific experiences [3].



	<img alt="" class="aligncenter size-full wp-image-1979" height="939" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture15.png" width="937" />▷&nbsp;Engram cells (pink) adjacent to non-engram cells (white). Source: Science



	&nbsp;



	It is generally believed that neuronal activity under external stimuli can induce the expression of immediate early genes in engram cells, providing the physical basis for engram formation.



	In 2015, a research team led by Professor Li-Huei Tsai from the Department of Brain and Cognitive Sciences at MIT used molecular and next-generation whole-genome sequencing methods to discover that neuronal activity stimulation can trigger the formation of DNA double-strand breaks (DSBs) in the promoters of immediate early genes (IEGs) such as Fos, Npas4, and Egr1. Even in the absence of external stimuli, targeted induction of DSBs in the promoter regions of IEGs is sufficient to induce their expression [4].



	The expression of immediate early genes is typically constrained by DNA topology, and DNA double-strand breaks can help rapidly relieve these constraints. Thus, DNA double-strand breaks play a crucial role in the expression of immediate early genes, learning, and memory.



	&nbsp;



	3. The History of DNA Double-Strand Breaks and Memory



	Maintaining DNA integrity is crucial for the faithful expression of genes. However, DNA is susceptible to various forms of damage during normal cellular functions. It is estimated that each cell may experience up to tens of thousands of DNA damage events daily, including base mismatches, oxidative base damage, single-strand breaks, and double-strand breaks (DSBs). All types of DNA damage can lead to adverse outcomes such as apoptosis or tumorigenesis. Among them, DSBs are considered the most cytotoxic as they pose a significant threat to the preservation of genetic and epigenetic information and can signal cell death. It is well-known that ionizing radiation-induced DSBs are a major pathogenic factor behind mutations, chromosomal aberrations, genetic instability, and carcinogenesis.



	However, not all DNA double-strand breaks are pathological. In healthy functioning cells, DSBs formed during mitosis are repaired through homologous recombination; in damaged or non-dividing cells, DSBs are primarily repaired through non-homologous end joining. These two mechanisms ensure the faithful and efficient repair of DSBs, playing a crucial adaptive role in the immune system and maintaining the delicate balance between damage pathology and adaptive immunity.



	In the central nervous system, the balance of DNA damage and repair in relation to memory appears to be more elusive. With aging and disease, DSBs increase. Studies show that DSBs in the hippocampus are considered detrimental to learning and memory throughout the lifespan. Mice repeatedly exposed to low-dose radiation exhibit persistent DSBs in the hippocampal dentate gyrus, leading to reduced neurogenesis and ultimately lower survival rates [5].



	At the same time, scientists have long suspected that DSBs might be involved in the memory process. After contextual fear memory retrieval, the active transcription marker histone H3 lysine 4 trimethylation (H3K4me3) increases in the hippocampus, which is one of the signs of DSBs. However, it is unclear how this epigenetic mark is regulated during reconsolidation [6].



	If DSBs are involved in learning and memory, then pharmacologically inducing DSBs would seem counterintuitive as it would impair memory. One possible reason is that an increase in DSBs without behavioral experience might be harmful and could inhibit DSBs triggered by behavioral experiences. Another possibility is that drug-induced and behavior-induced DSBs might jointly trigger DSBs in specific regions, whereas individual DSB events might not be sufficient to induce memory formation. Additionally, it is possible that some memory signals are associated with the repair of DSBs.



	These data suggest that DSBs have an adaptive role in the central nervous system.



	&nbsp;



	4. A New Cluster of Neurons, Distinct from Engram Cells



	So far, DNA double-strand breaks (DSBs) seem closely associated with engram cells, providing a pathway to explore how these cells function.



	Jovasevic&#39;s team discovered another group of neurons. They found that in a distinct neuronal population different from those expressing immediate early genes, DNA damage repair induced by fear memory became a source of long-term memory. DNA damage induced by memory stimuli could be linked to persistent cellular changes associated with long-term memory.



	Jovasevic&#39;s team conducted contextual fear conditioning training on mice and found that:



	(1) Four days after training the mice for fear conditioning (considered a short-term memory period), DNA and nuclear damage appeared in the hippocampus of the mice.



	(2) Observing the mice 21 days after fear memory stimulation (considered a long-term memory period), they found signs of inflammation in the hippocampus, activation of the Toll-like receptor 9 (TLR9) signaling pathway, and triggering of the cytosolic DNA innate immune response. This response is usually the body&#39;s defense mechanism against bacterial infection.



	<img alt="" class="aligncenter size-full wp-image-1978" height="533" src="https://admin.next-question.com/wp-content/uploads/2024/06/Picture16.png" width="830" />▷&nbsp;Mechanism of memory formation in inflammatory neurons and engram neurons. Source: Nature



	&nbsp;



	Although most DNA DSBs are rapidly repaired, those that persist show a specific centrosomal localization. The centrosome, known for its role in cell division, also plays an increasingly important role in the DNA damage response.



	Within the centrosome, DNA forms a crucial damage response complex, including the protein 53BP1, a key mediator of DSB repair. Studies show that removing Toll-like receptor 9 from hippocampal neurons prevents the accumulation of centrosomal DNA damage and DNA repair complexes, indicating that TLR9 signaling&#39;s critical function in neurons is maintaining genomic integrity.



	Jovasevic et al. further demonstrated how the transcription factor RELA localizes to the centrosome. These findings establish a molecular-level link between memory acquisition and its long-term reliable storage, providing substantial insights into the identification of these inflammatory neurons.



	So, how do these specific clusters of inflammatory neurons interact with engram cells? Researchers speculate that inflammatory neurons may play a crucial role in the stability and flexibility of memory, while engram neurons are responsible for triggering memory recall.



	Studies on engram neurons show that the baseline expression of the transcription factor CREB is the promoter of activity-induced gene expression in memory, initiating the recruitment of individual neurons into the engram. A similar mechanism may exist in the inflammatory cell population. DNA damage could cause some neurons to undergo persistent DNA damage and an innate immune response post-learning.



	The authors hypothesize that engram neurons might generate the initial memory signals, while inflammatory neurons support and shape the memory, helping it persist over time.



	&nbsp;



	5. Back to Semon: Memory and Inheritance



	In his book Die Mneme, Richard Semon proposed a broader unified theory emphasizing the similarities between memory and inheritance. This stemmed from his simple observation that parents&#39; learning and experiences could be transmitted to their offspring. This idea of &quot;soft inheritance&quot; originally came from Lamarck.



	Compared to the engram theory, Semon&#39;s hypothesis of intergenerational memory inheritance seemed more sensational. Once proposed, it was cast aside along with Lamarck&#39;s ideas, only to be picked up by the most optimistic of revivalists.



	In the 1950s and 60s, James V. McConnell, an experimental psychologist at the University of Michigan, and his colleagues used planarians to study memory processes, sparking a revival of the concept of soft inheritance of memory. He experimented with the regenerative capacity of planarians, bisecting a trained planarian and observing that both new worms retained some memory, demonstrated by their quicker completion of training tasks. Even more astonishing was that when trained planarians were ground up and fed to untrained ones, the latter also completed training tasks more readily, as if memory was transferred through material ingestion.



	These incredible experimental results sparked widespread controversy. A significant portion of the scientific community was skeptical, believing that similar conclusions could be reached with proper controls and accounting for inevitable observer bias.



	Due to prolonged scrutiny, McConnell became a target in the public eye, even receiving a letter bomb that resulted in his hearing loss.



	However, in 2013, a team led by biologists Tal Shomrat and Michael Levin from Tufts University in Massachusetts published a paper fundamentally supporting McConnell&#39;s findings. They used different floor textures and varying lighting as cues for adaptive memory, testing decapitated worms&#39; rapid adaptation to previous environments [7].



	That same year, a team led by Brian Dias from Emory University in the U.S. discovered the possibility of intergenerational inheritance in mammals. They found that mice could pass on an olfactory fear memory of acetophenone to their F1 generation. The idea of intergenerational memory inheritance, once discarded, was brought back into consideration [8].



	As previously mentioned, DNA damage gradually accumulates post-translational modifications of histones, which may be crucial for memory formation. Similar to epigenetic research, the repair of DNA double-strand breaks (DSBs) might be a vital component of transmitting information to offspring.



	Using chromatin immunoprecipitation (ChIP) analysis, researchers found significant overlap between DNA DSBs and histone H3 lysine 4 methylation (H3K4me3) in the CA1 region of the hippocampus during memory reconsolidation. Additionally, the phosphorylation of histone H2A (H2A.XpS139) and H3K4me3 levels were associated with long-term memory impairment, suggesting that DSB repair plays an indispensable role in the memory reconsolidation process. Moreover, the monoubiquitination of histone H2B at lysine 120 (H2Bub) is crucial for hippocampal memory formation [6].



	Overall, increased DSBs enhance epigenetically mediated transcriptional control, potentially representing a new mechanism for memory reconsolidation. The story of the engram seems to have returned once again to Semon.



	&nbsp;



	6. Outlook



	Of course, defining a new cluster of cells cannot rely on isolated findings alone. The discovery of engram cells caused a wave in the academic community a decade ago, supported by a robust triad of verification: observational studies, necessity through knockout studies, and sufficiency through knock-in studies.



	Regarding Jovasevic&#39;s research, determining how to identify these specific inflammatory neurons and distinguish them from other neuronal populations remains a current challenge. Especially considering that DNA damage plays a crucial role in aging and neurodegeneration, the emergence of these inflammatory neurons raises an urgent question: Toll-like receptor 9 (TLR9) activation typically involves microglia, the central nervous system&#39;s immune cells, which usually leads to neurodegeneration. So why, in neurons, does TLR9 activation contribute to memory formation instead?



	How can we differentiate between harmful DNA damage and inflammation and the immune response essential for memory?



	The activity or functional characteristics of this neuronal cluster in the processes of memory acquisition, consolidation, or retrieval remain unclear. Throughout the memory process, recording electrical activity and selectively manipulating inflammatory neurons are necessary to better understand these cells&#39; contribution to memory persistence.



	Answering these questions will require significant effort. Just as Dal&iacute; depicted memory and its instability in his iconic surrealist style, the neuronal cluster discovered by Jovasevic and colleagues still needs a comprehensive framework to support the persistence and disintegration of memory.



	&nbsp;

[:zh]


[:]

Memory, Originating from DNA Disintegration?

[:en]Memory, Originating from DNA Disintegration?[:]

As we navigate our environment, observing moving objects or experiencing shifting vistas as we move, we engage in a process known as visual motion. Numerous methods exist for studying visual motion, including electroencephalography (EEG). Traditionally, investigating visual motion with EEG requires stringent experimental conditions, which can limit its ecological validity.
In real-world scenarios, our brains integrate complex, dynamic, multimodal visual motion cues to guide the execution of movements. Researchers have identified the parieto-occipital cortex as a key brain region involved in controlling goal-directed movements such as reaching, grasping, catching, or avoiding objects. The fast-paced game of ping-pong, for example, demands rapid, full-body reactive integration of visual motion cues. To understand the neuronal activity in the parieto-occipital cortex during real-world ping-pong play, a research team from the University of Florida conducted a study using high-density scalp EEG for quantitative analysis of the electrodynamics in this area.
This research examined differences in power spectral density, event-related spectral perturbations, inter-trial phase coherence, event-related potentials, and event-related phase coherence in parieto-occipital source localization clusters. Subjects played ping-pong against both a robot and a human opponent. The results showed that playing against a robot led to greater θ-band power fluctuations before and after hitting the ball, stronger inter-phase consistency, more pronounced deflections in event-related potentials, and increased event-related coherence among parieto-occipital neural clusters, compared to playing against a human. The researchers concluded that robot-assisted motion training elicits distinct brain dynamics compared to human-assisted training, resulting in heightened brain engagement. These findings were published in eNeuro. 
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1537" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-1.png" alt="" width="456" height="178" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-1.png 456w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-1-300x117.png 300w" sizes="(max-width: 456px) 100vw, 456px" />
Figure 1: The cover of the research paper. Source: eNeuro.
&nbsp; 
Table tennis, a sport requiring reactive agility, offers a rich context for research into sensory-motor integration, anticipatory movement, and neural control mechanisms of object interception. The close proximity of opponents in a match necessitates rapid perception, planning, and striking within fractions of a second. Studies have shown that professional table tennis players, compared to non-athletes, exhibit quicker reflexes, faster hand movements, and more efficient visual-motor integration skills.
Additionally, table tennis has been found to mitigate cognitive decline, enhance executive functions, and sharpen mental acuity when used as a therapeutic intervention. Investigating table tennis&#8217; neural control mechanisms aids in understanding the neurophysiological principles behind these benefits.
In this research, 37 participants, all right-handed, with an average age of 23.5 years, normal vision, good health, and proficient table tennis skills, were involved. Each participant wore an EEG system equipped with 120 scalp electrodes during the experiments, with brain electrical activity recorded at 500 Hz. The table tennis rackets had wooden handles fitted with inertial measurement units for precise timing of ball-striking events. Video recordings were also made of all experimental sessions to verify the accuracy of the hit events.
The experiment was divided into four 15-minute intervals, with each interval consisting of 7.5 minutes of play against a human player and three consecutive 2.5-minute sessions against a robot. The robot sessions alternated between static and dynamic states. In static trials, the ball&#8217;s trajectory was predictable, requiring no movement from the participants to strike the ball. In dynamic trials, the robot varied the ball&#8217;s direction on the table, necessitating participant movement to hit the ball. 
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1542" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-3.jpg" alt="" width="757" height="488" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-3.jpg 757w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-3-300x193.jpg 300w" sizes="(max-width: 757px) 100vw, 757px" />
Figure 3: Cortical dipoles for all subjects. Axial view (left), sagittal view (center), and coronal view (right). Left frontal lobe (purple), right frontal lobe (magenta), left sensorimotor area (cyan), supplementary motor area (yellow), right sensorimotor area (dark red), left parieto-occipital area (light green), anterior cingulate area (blue), right parieto-occipital area (red), cingulate area (dark green). Source: From the paper.
&nbsp; 
The research compared brain electrical activity in the parieto-occipital area when subjects played against human players and robots. The power spectral density (PSD) graphs of the parieto-occipital clusters revealed both non-oscillatory and oscillatory components, with oscillatory power predominantly in the α and low beta bands for both human and robot opponents. Notably, the left parieto-occipital cortex showed higher α power during human trials than robot trials, while the anterior cingulate&#8217;s α power was higher during robot play. No significant differences were observed in the right parieto-occipital cortex and cingulate cortex between the two trial types. 
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1541" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-4.png" alt="" width="482" height="647" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-4.png 482w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-4-223x300.png 223w" sizes="(max-width: 482px) 100vw, 482px" />
Figure 4: Power spectra during play against human players (left, blue) and robots (center, magenta). The average power spectral density with non-periodic fitting by the FOOOF model (gray). The flattened spectrum (right), calculated by subtracting the non-periodic component, compares trials against human players (blue) and robots (magenta). Source: From the paper.
&nbsp; 
During human play, the left parieto-occipital cluster&#8217;s event-related spectral perturbations displayed desynchronization in the α and β bands after receiving the ball and synchronization before striking the ball. Robot play, in contrast, showed additional θ and α desynchronization before receiving the ball and increased phase consistency in these bands around 500ms after the robot served the ball. Compared to human play, robot play also exhibited stronger deflections and phase consistency in event-related potentials, with more pronounced phase-locked activity. 
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1540" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-5.png" alt="" width="592" height="629" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-5.png 592w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-5-282x300.png 282w" sizes="(max-width: 592px) 100vw, 592px" />
Figure 5: Comparison of time-frequency characteristics during play against human players and robots. Average results for the left parieto-occipital cluster group during trials against human players (left) and robots (center). Significant differences (right) are indicated in orange (p &lt; 0.05). Source: From the paper.
&nbsp; 
Moreover, event-related phase coherence was observed in the θ, α, and γ bands among the parieto-occipital clusters. θ coherence occurred at the appearance of the ball and about 250 ms afterward. Robot trials demonstrated higher event-related phase coherences than human trials, with statistically significant differences in coherence between the anterior cingulate and left/right parieto-occipital lobes. Specifically, the anterior cingulate&#8217;s θ coherence with the left parieto-occipital cluster was stronger in robot trials at the ball&#8217;s appearance, and with the right cluster, it was stronger both at the ball&#8217;s appearance and around 250ms afterward. 
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1539" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-6.png" alt="" width="919" height="944" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-6.png 919w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-6-292x300.png 292w, https://admin.next-question.com/wp-content/uploads/2024/02/PPP-6-768x789.png 768w" sizes="(max-width: 919px) 100vw, 919px" />
Figure 6: Phase consistency during trials against human players and robots. Comparing trials against human players (blue, lower diagonal) and robots (magenta, upper diagonal). Increases in event-related consistency significance are marked in red (p &lt; 0.05). Source: From the paper.
&nbsp; 
In summary, the research revealed distinct dynamics in the parieto-occipital cortex&#8217;s neuronal activity related to table tennis strokes, with notable differences between playing against humans and robots. Key differences included greater θ power fluctuations in the brain during robot play at stroke events and higher event-related coherence among parieto-occipital clusters. The increased brain activity while waiting for the robot to serve may stem from the absence of human-like cues, such as racket swings, requiring players to stay fully alert to respond quickly to the robot&#8217;s unpredictable strokes.
The research underscores that our brains process these two experiences very differently, suggesting that training with machines might not replicate the experience of playing against real opponents.
&nbsp; 
Author&#8217;s self-evaluation:
Daniel P. Ferris:
Robots are increasingly becoming integral to our daily lives. Some companies are building robots designed to interact with humans, while others are focused on developing social assistive robots to help the elderly. The interaction between humans and robots is distinct from that of human-to-human interaction. Our long-term goal is to understand how the brain responds to these differences.
This research reveals that when competing against another person, players&#8217; neurons tend to synchronize their activity. However, when facing a robotic opponent, these neurons tend to desynchronize. Imagine a football stadium with 100,000 people cheering in unison; this represents brain synchronization, indicating a relaxed state of the brain. Conversely, if these 100,000 people were all chatting with friends, their activities would be unsynchronized, akin to brain desynchronization. Brain desynchronization often suggests that the brain is engaged in extensive computations.
Amanda Studnicki:
Table tennis, a sport that involves multiple senses like vision, balance (vestibular sense), and hearing, provides crucial information to the central brain for creating motor plans. While researchers have studied simple movements such as reaching and grasping in static tasks, our aim is to understand how the brain processes more complex actions, like tracking and intercepting a moving ball. Table tennis, therefore, serves as an ideal subject for study. Training with robots, we believe, is incredibly valuable. As sports robots continue to evolve over the next 10 to 20 years, we anticipate seeing an increased use of robots in player training.

[:en]As we navigate our environment, observing moving objects or experiencing shifting vistas as we move, we engage in a process known as visual motion. Numerous methods exist for studying visual motion, including electroencephalography (EEG). Traditionally, investigating visual motion with EEG requires stringent experimental conditions, which can limit its ecological validity.

In real-world scenarios, our brains integrate complex, dynamic, multimodal visual motion cues to guide the execution of movements. Researchers have identified the parieto-occipital cortex as a key brain region involved in controlling goal-directed movements such as reaching, grasping, catching, or avoiding objects. The fast-paced game of ping-pong, for example, demands rapid, full-body reactive integration of visual motion cues. To understand the neuronal activity in the parieto-occipital cortex during real-world ping-pong play, a research team from the University of Florida conducted a study using high-density scalp EEG for quantitative analysis of the electrodynamics in this area.

This research examined differences in power spectral density, event-related spectral perturbations, inter-trial phase coherence, event-related potentials, and event-related phase coherence in parieto-occipital source localization clusters. Subjects played ping-pong against both a robot and a human opponent. The results showed that playing against a robot led to greater θ-band power fluctuations before and after hitting the ball, stronger inter-phase consistency, more pronounced deflections in event-related potentials, and increased event-related coherence among parieto-occipital neural clusters, compared to playing against a human. The researchers concluded that robot-assisted motion training elicits distinct brain dynamics compared to human-assisted training, resulting in heightened brain engagement. These findings were published in eNeuro.
&nbsp;
<img class="aligncenter size-full wp-image-1537" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-1.png" alt="" width="456" height="178" />
Figure 1: The cover of the research paper. Source: eNeuro.
&nbsp;
Table tennis, a sport requiring reactive agility, offers a rich context for research into sensory-motor integration, anticipatory movement, and neural control mechanisms of object interception. The close proximity of opponents in a match necessitates rapid perception, planning, and striking within fractions of a second. Studies have shown that professional table tennis players, compared to non-athletes, exhibit quicker reflexes, faster hand movements, and more efficient visual-motor integration skills.

Additionally, table tennis has been found to mitigate cognitive decline, enhance executive functions, and sharpen mental acuity when used as a therapeutic intervention. Investigating table tennis' neural control mechanisms aids in understanding the neurophysiological principles behind these benefits.

In this research, 37 participants, all right-handed, with an average age of 23.5 years, normal vision, good health, and proficient table tennis skills, were involved. Each participant wore an EEG system equipped with 120 scalp electrodes during the experiments, with brain electrical activity recorded at 500 Hz. The table tennis rackets had wooden handles fitted with inertial measurement units for precise timing of ball-striking events. Video recordings were also made of all experimental sessions to verify the accuracy of the hit events.

The experiment was divided into four 15-minute intervals, with each interval consisting of 7.5 minutes of play against a human player and three consecutive 2.5-minute sessions against a robot. The robot sessions alternated between static and dynamic states. In static trials, the ball's trajectory was predictable, requiring no movement from the participants to strike the ball. In dynamic trials, the robot varied the ball's direction on the table, necessitating participant movement to hit the ball.
&nbsp;
<img class="aligncenter size-full wp-image-1542" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-3.jpg" alt="" width="757" height="488" />
Figure 3: Cortical dipoles for all subjects. Axial view (left), sagittal view (center), and coronal view (right). Left frontal lobe (purple), right frontal lobe (magenta), left sensorimotor area (cyan), supplementary motor area (yellow), right sensorimotor area (dark red), left parieto-occipital area (light green), anterior cingulate area (blue), right parieto-occipital area (red), cingulate area (dark green). Source: From the paper.
&nbsp;
The research compared brain electrical activity in the parieto-occipital area when subjects played against human players and robots. The power spectral density (PSD) graphs of the parieto-occipital clusters revealed both non-oscillatory and oscillatory components, with oscillatory power predominantly in the α and low beta bands for both human and robot opponents. Notably, the left parieto-occipital cortex showed higher α power during human trials than robot trials, while the anterior cingulate's α power was higher during robot play. No significant differences were observed in the right parieto-occipital cortex and cingulate cortex between the two trial types.
&nbsp;
<img class="aligncenter size-full wp-image-1541" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-4.png" alt="" width="482" height="647" />
Figure 4: Power spectra during play against human players (left, blue) and robots (center, magenta). The average power spectral density with non-periodic fitting by the FOOOF model (gray). The flattened spectrum (right), calculated by subtracting the non-periodic component, compares trials against human players (blue) and robots (magenta). Source: From the paper.
&nbsp;
During human play, the left parieto-occipital cluster's event-related spectral perturbations displayed desynchronization in the α and β bands after receiving the ball and synchronization before striking the ball. Robot play, in contrast, showed additional θ and α desynchronization before receiving the ball and increased phase consistency in these bands around 500ms after the robot served the ball. Compared to human play, robot play also exhibited stronger deflections and phase consistency in event-related potentials, with more pronounced phase-locked activity.
&nbsp;
<img class="aligncenter size-full wp-image-1540" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-5.png" alt="" width="592" height="629" />
Figure 5: Comparison of time-frequency characteristics during play against human players and robots. Average results for the left parieto-occipital cluster group during trials against human players (left) and robots (center). Significant differences (right) are indicated in orange (p &lt; 0.05). Source: From the paper.
&nbsp;
Moreover, event-related phase coherence was observed in the θ, α, and γ bands among the parieto-occipital clusters. θ coherence occurred at the appearance of the ball and about 250 ms afterward. Robot trials demonstrated higher event-related phase coherences than human trials, with statistically significant differences in coherence between the anterior cingulate and left/right parieto-occipital lobes. Specifically, the anterior cingulate's θ coherence with the left parieto-occipital cluster was stronger in robot trials at the ball's appearance, and with the right cluster, it was stronger both at the ball's appearance and around 250ms afterward.
&nbsp;
<img class="aligncenter size-full wp-image-1539" src="https://admin.next-question.com/wp-content/uploads/2024/02/PPP-6.png" alt="" width="919" height="944" />
Figure 6: Phase consistency during trials against human players and robots. Comparing trials against human players (blue, lower diagonal) and robots (magenta, upper diagonal). Increases in event-related consistency significance are marked in red (p &lt; 0.05). Source: From the paper.
&nbsp;
In summary, the research revealed distinct dynamics in the parieto-occipital cortex's neuronal activity related to table tennis strokes, with notable differences between playing against humans and robots. Key differences included greater θ power fluctuations in the brain during robot play at stroke events and higher event-related coherence among parieto-occipital clusters. The increased brain activity while waiting for the robot to serve may stem from the absence of human-like cues, such as racket swings, requiring players to stay fully alert to respond quickly to the robot's unpredictable strokes.

The research underscores that our brains process these two experiences very differently, suggesting that training with machines might not replicate the experience of playing against real opponents.

&nbsp;
Author's self-evaluation:

Daniel P. Ferris:

Robots are increasingly becoming integral to our daily lives. Some companies are building robots designed to interact with humans, while others are focused on developing social assistive robots to help the elderly. The interaction between humans and robots is distinct from that of human-to-human interaction. Our long-term goal is to understand how the brain responds to these differences.

This research reveals that when competing against another person, players' neurons tend to synchronize their activity. However, when facing a robotic opponent, these neurons tend to desynchronize. Imagine a football stadium with 100,000 people cheering in unison; this represents brain synchronization, indicating a relaxed state of the brain. Conversely, if these 100,000 people were all chatting with friends, their activities would be unsynchronized, akin to brain desynchronization. Brain desynchronization often suggests that the brain is engaged in extensive computations.

Amanda Studnicki:

Table tennis, a sport that involves multiple senses like vision, balance (vestibular sense), and hearing, provides crucial information to the central brain for creating motor plans. While researchers have studied simple movements such as reaching and grasping in static tasks, our aim is to understand how the brain processes more complex actions, like tracking and intercepting a moving ball. Table tennis, therefore, serves as an ideal subject for study. Training with robots, we believe, is incredibly valuable. As sports robots continue to evolve over the next 10 to 20 years, we anticipate seeing an increased use of robots in player training.[:]

Playing Ping-Pong with Robots: Does It Excite Our Brains More

[:en]Playing Ping-Pong with Robots: Does It Excite Our Brains More[:]

New Frontiers

Taking advantage of intracranial electroencephalographic (iEEG) recordings using depth and subdural electrodes from 36 human subjects using the same task, Dr. Gerwin Schalk, Director of the Chen Frontier Lab for Applied Neurotechnology, and his colleagues applied single-trial and cross-trial analyses to high-frequency iEEG activity. The results, published in the journal Cerebral Cortex, show that the neural activation was widely distributed across the human brain both within it and on the surface, and focused specifically on certain areas in the parietal, frontal, and occipital lobes, where parietal lobes present significant left lateralization on the activation. The team was also able to demonstrate temporal differences across these brain regions.
&nbsp;
Intracranial electroencephalography (iEEG), recorded using either subdural electrodes (ECoG) or deep electrodes (SEEG), is capable of recording neural activity over a fairly wide range of brain regions, with millimeter-level spatial resolution, millisecond-level temporal resolution. It is currently used for the preoperative detection of neural activity in patients with tumors or intractable epilepsy. iEEG recordings also have the ability to capture broadband γ-wave signals (frequencies &gt; 60 Hz), which is a reliable indicator of local neuronal activity.
&nbsp;
The current study shows that neurons from a wide range of regions are involved in motor tasks based on visual cues. And they also observed that the processing of motor tasks based on visual cues requires the engagement of neural networks across both cortical and subcortical regions.
&nbsp;
In addition to this, they found a significant left lateralization of neuronal activity in the parietal lobe in a visually cued motor task. At the same time, it has been shown that neuronal activation of this left lateralization tendency is independent of handedness. These observations suggest that there is a movement observation/execution network involving this region and that this network may mediate human recognition of the targets of observed movements. The current study extends the results to the whole brain and provides a relatively complete overview of the “footprint” of neural activity during the task.
&nbsp;
This study demonstrates the overall large-scale spatio-temporal neural activity evolution of the human brain during a visually cued motor task and evaluates the possible functions of different brain regions. The findings provide implications for research on motor-related brain-computer interfaces: decoding motor parameters requires more effort into brain regions such as the frontal and parietal lobes in addition to traditional sensorimotor regions.
&nbsp;
<a href="https://academic.oup.com/cercor/article-abstract/33/17/9764/7225904?redirectedFrom=fulltext&amp;login=false">Read the paper on Cerebral Cortex</a>

[:en]Taking advantage of intracranial electroencephalographic (iEEG) recordings using depth and subdural electrodes from 36 human subjects using the same task, Dr. Gerwin Schalk, Director of the Chen Frontier Lab for Applied Neurotechnology, and his colleagues applied single-trial and cross-trial analyses to high-frequency iEEG activity. The results, published in the journal Cerebral Cortex, show that the neural activation was widely distributed across the human brain both within it and on the surface, and focused specifically on certain areas in the parietal, frontal, and occipital lobes, where parietal lobes present significant left lateralization on the activation. The team was also able to demonstrate temporal differences across these brain regions.

&nbsp;

Intracranial electroencephalography (iEEG), recorded using either subdural electrodes (ECoG) or deep electrodes (SEEG), is capable of recording neural activity over a fairly wide range of brain regions, with millimeter-level spatial resolution, millisecond-level temporal resolution. It is currently used for the preoperative detection of neural activity in patients with tumors or intractable epilepsy. iEEG recordings also have the ability to capture broadband γ-wave signals (frequencies &gt; 60 Hz), which is a reliable indicator of local neuronal activity.

&nbsp;

The current study shows that neurons from a wide range of regions are involved in motor tasks based on visual cues. And they also observed that the processing of motor tasks based on visual cues requires the engagement of neural networks across both cortical and subcortical regions.

&nbsp;

In addition to this, they found a significant left lateralization of neuronal activity in the parietal lobe in a visually cued motor task. At the same time, it has been shown that neuronal activation of this left lateralization tendency is independent of handedness. These observations suggest that there is a movement observation/execution network involving this region and that this network may mediate human recognition of the targets of observed movements. The current study extends the results to the whole brain and provides a relatively complete overview of the “footprint” of neural activity during the task.

&nbsp;

This study demonstrates the overall large-scale spatio-temporal neural activity evolution of the human brain during a visually cued motor task and evaluates the possible functions of different brain regions. The findings provide implications for research on motor-related brain-computer interfaces: decoding motor parameters requires more effort into brain regions such as the frontal and parietal lobes in addition to traditional sensorimotor regions.

&nbsp;

<a href="https://academic.oup.com/cercor/article-abstract/33/17/9764/7225904?redirectedFrom=fulltext&amp;login=false">Read the paper on Cerebral Cortex</a>[:zh]穿越十字路口时，行人会秉承“红灯停，绿灯行”的原则，根据交通信号灯的颜色，决定接下来的行为。在日常生活中，对视觉线索做出反应是人们最基础最重要的活动。这个过程听起来简单，但其中涉及的神经元时空动态变化却非常复杂。

&nbsp;

在时空尺度上揭示这个过程中大脑动态变化的机制对于人类神经科学的研究以及潜在的应用都有极其重要的作用。

&nbsp;

近日，天桥脑科学研究院（Tianqiao and Chrissy Chen Institute，TCCI）应用神经技术前沿实验室主任Gerwin Schalk教授与国内多所高校的研究团队联合在Cerebral Cortex发表题为“Spatio-temporal evolution of human neural activity during visually cued hand movements”的研究。他们利用深部电极或颅下电极记录了36名受试者面对相同任务时颅下脑电活动，并对高频脑电活动进行了分析。其结果表明人类神经元的激活广泛分布在整个大脑，集中于顶、额叶以及枕叶的特定区域，其中顶叶激活出现了明显的左侧化。并且，各个脑区的神经元激活呈现明显的时间差异。

&nbsp;

[caption id="attachment_762" align="aligncenter" width="600"]<img class="wp-image-762" src="https://admin.next-question.com/wp-content/uploads/2023/11/62.png" alt="" width="600" height="213" /> ▷图注：文章截图。图源：Cerebral Cortex官网[/caption]

&nbsp;
<h3>iEEG脱颖而出</h3>
&nbsp;

在常用的神经成像技术中，功能磁共振成像技术（fMRI）具有很好的空间分辨率，能够在大脑尺度上辨别功能活跃网络，但该技术是基于血氧水平的变化，无法反映不同脑区快速变化的神经活动，时间分辨率低。脑电图和脑磁图等非侵入性电生理方法则相反，它们能在毫秒级别研究大脑神经活动的变化，空间分辨率却受限。

&nbsp;

而颅内脑电图（iEEG）使用硬膜下电极（ECoG）或深部电极（SEEG）进行记录，能够记录相当广泛的脑区的神经活动，其具有毫米级的空间分辨率、毫秒级的时间分辨率，目前被用于肿瘤病人或难治性癫痫病人的神经活动的术前检测。iEEG记录还具有捕获宽带γ波信号(频率&gt;60 Hz的信号)的能力，该信号是局部神经元活动的可靠指标。

&nbsp;

基于这些特性，利用iEEG信号可以绘制记录位点间潜在的任务相关神经元的时空演化图谱。但这种记录方法也并非没有缺点——其局限性是存在稀疏采样。不过，通过对多个受试者进行记录，获得更多的数据点，并针对他们在相同任务中的数据进行组合分析，就可以增加对脑动态的把握、减轻这种限制，从而得出关于人类大脑在常见行为中的大规模时空动态特征的信息。

&nbsp;
<h3>基于视觉线索的运动任务</h3>
&nbsp;

研究者采集了36位右利手难治性癫痫患者（女14例，男22例，年龄26.0±6.2岁）的iEEG数据。这些患者均接受过用于评估癫痫病灶的电极植入，其中34位患者植入的是SEEG电极，2位患者植入的是ECoG电极。

&nbsp;

iEEG数据采集过程如下：

&nbsp;

每次试验过程中，当LCD屏幕出现“rest”指示时，受试者在接下来4秒内不能有任何活动；在视觉线索出现的前1秒，屏幕会显示提醒，意味着即将出现手势图片。屏幕出现视觉线索时，受试者被要求尽可能快地做出和图片相同的动作，直到5秒后图片消失。对于每一位受试者，研究人员收集了100次试验的iEEG数据。

&nbsp;

[caption id="attachment_763" align="aligncenter" width="600"]<img class="wp-image-763" src="https://admin.next-question.com/wp-content/uploads/2023/11/63.png" alt="" width="600" height="569" /> ▷图注：基于视觉线索的运动任务。图源：论文补充材料[/caption]

&nbsp;
<h3>大脑神经活动的分布特征及时空演化</h3>
&nbsp;

此次研究表明，相当广泛区域的神经元都参与基于视觉线索下的运动任务。除了最活跃的中央回、顶叶、枕叶、额叶和颞叶，在岛叶、海马旁回和海马等脑深部结构和边缘系统中也观察到神经元的活动。并且，他们还观察到，处理基于视觉线索的运动任务需要跨越皮层和皮层下两个部分募集神经网络。

&nbsp;

[caption id="attachment_764" align="aligncenter" width="600"]<img class="wp-image-764" src="https://admin.next-question.com/wp-content/uploads/2023/11/64.png" alt="" width="600" height="364" /> ▷图注：基于视觉线索的运动任务中神经活动的时空分布。图源：论文[/caption]

&nbsp;

除此之外，他们还发现，在基于视觉线索的运动任务中，顶叶的神经元活动存在显著的左侧化倾向。其他研究者在工具-行为观察、听觉和视觉刺激处理、手部形状编码的视觉和运动想象等认知过程中，也观察到类似的现象。同时，有研究表明这种左侧化倾向的神经元激活与利手性无关。

&nbsp;

这些观察结果表明，存在一个涉及该区域的动作观察/执行网络，而且该网络可能介导了人类对观察到的动作的目标的识别。

&nbsp;

紧接着，他们分析了该任务中神经活动的时空演化，发现神经活动首先出现在枕叶外侧部分，顶叶的上部和后部，以及颞叶的后部和下部，随后向前额叶扩散，最终在中央回结束。

&nbsp;

先前的研究也得到了与此一致的结果，但本次研究将结果扩展到了更广泛的整个大脑区域，因此可以相对完整地概述任务期间神经活动的“足迹”。

&nbsp;
<h3>特定脑区的功能</h3>
&nbsp;

研究者根据神经活动开始时间与刺激开始时间/反应启动时间之间的联系，分析了各脑区神经活动所代表的意义：“感觉”或“运动”。

&nbsp;

在完成基于视觉线索的运动任务过程中，早期激活的神经元与视觉刺激传递更相关，而晚期激活的神经元则与运动反应更相关。

&nbsp;

[caption id="attachment_765" align="aligncenter" width="600"]<img class="wp-image-765" src="https://admin.next-question.com/wp-content/uploads/2023/11/65.png" alt="" width="600" height="498" /> ▷图注：大脑不同区域激活与基于视觉线索的任务之间的评估结果。图源：论文[/caption]

&nbsp;

通过神经元激活发生的时间和比例，研究人员发现，最早激活的侧枕叶皮层在视觉信息处理中至关重要作用。然后，顶叶和颞叶也有较高比例的激活，表明它们在视觉信息处理中发挥重要作用。随着神经活动过程蔓延到额叶和中央回，这种视觉表征逐渐减弱。此外，顶叶和额叶在早期阶段明显参与了与运动反应启动相关的神经活动。这些发现表明，在将感觉信息转化为运动相关信息时，顶叶和额叶发挥重要作用：其中顶叶参与了转化过程的早期阶段，而额叶更倾向于参与运动执行阶段。

&nbsp;

而在“运动”过程的最后阶段，中央回呈现出最高比例的与反应锁定的神经活动，表明它们在运动执行和体感处理中发挥作用。不仅如此，中央前区中还检测到少数与早期刺激锁定的神经活动。这些发现使人们对这一运动区域的功能有了更全面的理解：中央前回是线索-行动网络的组成部分，以便对环境刺激做出即时反应。总之，在运动执行的最后阶段，中央前回接受高度加工的感觉输入和其他内部信号而产生运动信号，而运动执行后，中央后回中产生体感反馈。

&nbsp;
<h3>总结</h3>
&nbsp;

该研究展示了在基于视觉线索的运动任务期间，人脑整体大规模时空神经活动演化，并评估了不同脑区可能的功能。值得注意的是，研究仅集中于神经记录的高频成分，而未涉及一些与运动相关的低频活动。

&nbsp;

但是，此次发现不仅加深了人们对基于视觉任务下神经活动的理解，还为与运动相关的脑机接口研究提供一定借鉴作用：解码运动参数除了需要关注传统的感觉运动区域外，还需关注额叶和顶叶等脑区。

&nbsp;

&nbsp;

&nbsp;

<a href="https://academic.oup.com/cercor/article-abstract/33/17/9764/7225904?redirectedFrom=fulltext&amp;login=false">阅读原文</a>[:]

How do neurons coordinate with each other to generate hand movements?

[:en]How do neurons coordinate with each other to generate hand movements?[:zh]神经元们如何协作做出手势动作？[:]

They took the road less traveled by to unravel the most complicated mathematical mysteries

[:en]They took the road less traveled by to unravel the most complicated mathematical mysteries[:zh]他们自愿钻进学术“天坑”，只为解开这道最复杂的数学题｜追问观察[:]

[:en]
	For readers with a background in high school mathematics or statistics, the most recognized probability distribution is the normal distribution, which describes how an independent variable behaves under the influence of random factors after repeated trials. However, the log-normal distribution, characterized by its long tail, is more commonly observed in the brain.



	What biological mechanisms give rise to the log-normal distribution? How does the brain utilize this distribution across different scales&mdash;from cellular to perceptual&mdash;to support its complex functions? How does the brain&#39;s adherence to this distribution pattern enable efficient computation and inspire advancements in artificial intelligence? This article will review foundational theories and several subsequent studies in an attempt to answer these questions.


<h3>
	1. Normal Distribution vs. Log-Normal Distribution
</h3>


	What is a log-normal distribution? To grasp this concept, imagine participating in a virtual stock market simulation game. At the start of the game, each player has the same initial capital. Players must buy and sell stocks daily based on random market fluctuations. The decisions and luck of each player determine their gains or losses. On the first day, the distribution of players&#39; gains and losses will follow a normal distribution, with a few players experiencing significant losses, a few others achieving substantial gains, and most falling somewhere in between with minor gains or losses. 
	However, over time, you&#39;ll notice that although the daily fluctuations are random (with each decision&rsquo;s outcome following a normal distribution), the final distribution of each player&rsquo;s capital becomes skewed and long-tailed. This is the log-normal distribution. 
	Because the player&rsquo;s capital changes based on their previous balance, a cumulative effect takes place. As a result, most players&rsquo; capital tends to cluster at lower levels. Those who consistently break even or incur small losses over time will see their capital stabilize at lower amounts. Conversely, a very small number of players will accumulate significant wealth as consistent winners continue to gain more, resulting in a long-tailed distribution*.



	<img alt="" class="aligncenter size-full wp-image-2292" height="424" src="https://admin.next-question.com/wp-content/uploads/2024/08/对数-01.png" width="605" />



	▷ Fig.1. A comparison of normal distribution (green) and log-normal distribution (red). The compares the normal distribution (green) and the log-normal distribution (red). In the stock market example, the horizontal axis represents capital, and the vertical axis represents the number of players. In the case of neurons, the horizontal axis represents activation frequency, and the vertical axis represents the number of neurons. In the stock market example, because each player&#39;s capital changes by a fixed percentage (e.g., increasing or decreasing by 50%) based on their previous balance with each trade, only a very small portion of players consistently lose (e.g., from 100 to 50 to 25 to 12.5 to 6.25), while a few consistently win (e.g., from 100 to 150 to 225 to 337.5). Most players fall somewhere in between, fluctuating between gains and losses (e.g., 100 to 150 to 75 to 112.5 to 56.25; 100 to 50 to 75 to 37.5 to 56.25), with the majority clustering at lower values.



	&nbsp;



	In academic circles, random variables following a log-normal distribution are often described using terms like power law or scale-free, all referring to the same underlying phenomenon.


<h3>
	2. Log-Normal Distributions in Neuronal Activations, Dendritic Spine Sizes, and Synaptic Densities
</h3>


	In the brain, the behavior of neurons is analogous to the lights in a city. Most of the lights remain off, but occasionally, a small number turn on, illuminating the surrounding area. Multiple studies have demonstrated that these intervals of &#39;illumination,&#39; reflected in the firing rates of neurons, follow a log-normal distribution pattern [1]. This pattern is consistent across different species, including mice and humans. It is observed in both excitatory pyramidal neurons and inhibitory interneurons and is evident in brain regions ranging from the evolutionarily primitive cerebellum to the neocortex, which is responsible for higher cognitive functions.



	The asymmetry of the log-normal distribution means that a small portion of neurons in the brain are frequently active, while the majority remain relatively idle, only occasionally activating. Moreover, the amplitude of neuronal firing spikes and the weights of synapses also follow a log-normal distribution [1].



	&nbsp;



	<img alt="" class="aligncenter size-large wp-image-2293" height="663" src="https://admin.next-question.com/wp-content/uploads/2024/08/对数-02-1024x663.webp" width="1024" /> 
	▷ Fig.2. In various brain regions and environments, the firing rates of common types of neurons follow a log-normal distribution (both the horizontal and vertical axes have been logarithmically transformed, resulting in a linear scatter plot). Source: Reference 1.



	&nbsp;



	However, neurons do not simply &#39;turn on&#39; or &#39;off.&#39; Closer examination of their structure reveals spiny protrusions on dendritic branches known as dendritic spines. A single neuron may contain hundreds of dendritic spines, varying in size. These spines undergo plasticity during neuronal development and play critical roles in information storage and computation.



	A study observed that in the auditory cortex of mice, the sizes of dendritic spines also follow a log-normal distribution (Fig.3). Furthermore, the magnitude of their changes is proportional to their original size, resembling the fluctuations in stock market capital: the larger the investment, the greater the fluctuation.



	<img alt="" class="aligncenter size-large wp-image-2294" height="818" src="https://admin.next-question.com/wp-content/uploads/2024/08/04-1024x818.webp" width="1024" /> 
	▷ Fig.3. The sizes of dendritic spines follow a log-normal distribution. A: Schematic of a dendritic spine, B: Raw probability distribution, C, D: Log-transformed number of dendritic spines and probability distribution on the X-axis.



	&nbsp;



	The above analysis is based on recordings from individual neurons. However, with advances in connectomics, researchers can now map networks composed of neurons within specific regions. For example, a study of nine mammalian species [3] found that in seven of these species, neuronal density (the number of neurons per cubic millimeter) in the cerebral cortex follows a log-normal distribution (Fig.4). This suggests that the log-normal distribution is a fundamental principle of the macroscopic organization of the nervous system and should be considered a constraint when designing neuromorphic hardware.



	<img alt="" class="aligncenter size-full wp-image-2296" height="1020" src="https://admin.next-question.com/wp-content/uploads/2024/08/05.webp" width="1002" /> 
	▷ Fig.4. Distribution of neuronal densities across nine mammalian species. The degree of deviation is shown, with the two species that do not follow the log-normal distribution (GALAGO and OWL MONKEY) possibly due to uneven sampling and insufficient sample size. Source: Reference 3.



	&nbsp;


<h3>
	3. Is It True That We Only Use 10% of Our Brain?
</h3>


	The popular myth that we only use 10% of our brain arises from a fundamental misunderstanding of how brain activity is distributed. Brain activity is not uniformly distributed; rather, it exhibits a &#39;long-tail&#39; characteristic, with a small number of neurons being highly active while the majority are less active. The distribution of neuronal firing rates is often attributed to the complex interconnectedness of neurons [1].


<h4>
	(1) Interactions Between Neurons
</h4>


	Let&#39;s use the analogy of a large symphony orchestra to better understand this concept. In a symphony orchestra, each musician has their own instrument and plays at their own rhythm. During a performance, not all musicians play simultaneously. Instead, depending on the composition, some musicians play more actively, creating the music&#39;s climax, while others may rest, waiting for their next cue. This uneven activity pattern mirrors the activation states of neurons in the brain: most neurons are relatively &quot;quiet&quot; most of the time, but a small, ever-changing group of active neurons drives the brain&#39;s complex thinking and perception tasks.



	The scientific basis for this activity pattern is that interactions between neurons are not merely additive but cumulative, much like the harmonious coordination and mutual stimulation among musicians, involving complex regulation and feedback mechanisms. For example, in certain situations, the activity of one neuron may prompt other closely connected neurons to become active, forming a &quot;hotspot&quot; of activity. This phenomenon across the entire neural network results in a log-normal distribution pattern.



	In addition to explaining cumulative effects through neuronal connectivity, some studies use the &quot;Integrate-and-Fire Model&quot; to demonstrate that even individual neurons can exhibit a log-normal distribution in firing rates due to internal cumulative adjustments (such as the compounding errors in membrane potential discharge) [4].


<h4>
	(2) Hebbian Theory
</h4>


	Hebbian theory, which describes the self-organization of neurons, offers another explanation for the log-normal distribution observed in the nervous system [5]. Similar to relationships in a social network, neurons that frequently activate together form stronger connections. Whether in simple organisms like C. elegans or more complex species like fruit flies and mice, the number and density of neuronal connections exhibit a scale-free, long-tail characteristic&mdash;most connections are weak, while a few are very strong.



	The theoretical predictions of these models align with experimental results. Enhancing connections between co-activated neurons&mdash;including direct first-order connections and second- and third-order indirect connections (Fig.5a)&mdash;reproduces the actual distribution of connection strengths observed (solid line in Fig.5b). This suggests that the modularity of the nervous system arises from the generative mechanisms of the network rather than the specificity of species or neurons, further explaining the ubiquity of log-normal probability distributions.



	<img alt="" class="aligncenter size-large wp-image-2295" height="469" src="https://admin.next-question.com/wp-content/uploads/2024/08/06-1024x469.webp" width="1024" /> 
	▷ Fig.5. Hebbian theory-based explanation of the scale-free nature of neural activity.



	&nbsp;


<h4>
	(3) Energy Consumption and Evolutionary Perspective
</h4>


	From the perspective of reducing brain energy consumption and signal transmission distance, the log-normal distribution is advantageous [6]. Specifically, this is achieved by clustering a small number of neurons in a certain area to form a modular &quot;rich club&quot; and having a few neurons frequently activate. Therefore, modularity and the frequent activation of a small number of computational units should be guiding principles for designing neuromorphic computing systems.



	Evolutionarily, a ready group of neurons with dense connections and rapid activation can be seen as a core system within the brain, always prepared to act swiftly, which is crucial for species survival. However, the full functionality of the brain depends on another system provided by the majority of neurons with sparse connections and infrequent activation. These neurons participate less in daily activities but are crucial when adapting to new or uncommon stimuli. 
	Thus, instead of saying we use only 10% of our brain, we should understand that only 10% of the brain is actively working at any given moment. This reflects an optimal balance achieved by the brain to handle both common and uncommon stimuli across multiple timescales, a characteristic of the brain&#39;s criticality [7].


<h4>
	(4) Other Explanations
</h4>


	In addition to the models mentioned above, another explanation for the log-normal distribution in the nervous system [8] considers two different types of neurons. This model divides neurons into two categories: one that remains silent due to excessive &quot;silent&quot; signals (inhibitory neuron influence), and another that becomes highly active due to &quot;volume increase&quot; (excitatory neuron influence). The model assumes that the initial signals received by all neurons are uniformly distributed, like evenly distributed sounds. However, as these signals accumulate within neurons, those more influenced by excitatory signals amplify them, resulting in their activity output following a log-normal distribution.



	This model not only explains how the balance between excitation and inhibition shapes neural responses but also clarifies how the nervous system maintains sensitivity to small signals while avoiding overreaction to excessively strong inputs.



	&nbsp;



	<img alt="" class="aligncenter size-full wp-image-2297" height="290" src="https://admin.next-question.com/wp-content/uploads/2024/08/07.webp" width="384" /> 
	▷ Fig.6. Log-normal distribution among neurons due to nonlinear gain effects.



	&nbsp;



	Moreover, a model proposed by Per Bak in his book How Nature Works offers another perspective, known as the &quot;critical sandpile model.&quot; In this model, the accumulation and collapse of sand grains resemble the dynamic changes in neural networks: as the sandpile grows, small collapses occur frequently, while large collapses, requiring more accumulated sand, are rarer.



	Scientists, through studying the physical and mechanical properties of the cytoskeleton [9], have discovered that the aggregation and fragmentation behaviors of neurons resemble the collapse of a sandpile. This process leads to the distribution of neuronal density also exhibiting a log-normal distribution, which explains the uneven phenomenon in neural networks where a &quot;few neurons are highly active while the majority remain silent.&quot;



	<img alt="" class="aligncenter size-large wp-image-2299" height="1024" src="https://admin.next-question.com/wp-content/uploads/2024/08/nature-692x1024.webp" width="692" /> 
	▷ How Nature Works by Per Bak.



	&nbsp;


<h3>
	4. Conclusion: What Is the Value of Examining the Statistical Characteristics of the Nervous System?
</h3>


	Thoroughly analyzing the statistical characteristics of the nervous system allows researchers to uncover universal principles that span different species and brain regions, while also exploring how these features change under various environmental stimuli.



	For example, a study on rhesus monkeys found that neuronal firing rates follow a log-normal distribution when exposed to both familiar and unfamiliar visual stimuli [10]. However, the statistical properties of these firing rates differ depending on the familiarity of the environment. When viewing unfamiliar images, the long-tail feature of the distribution becomes more pronounced, with an increase in variance (Fig.8). This suggests that neurons become more actively engaged when processing novel stimuli, while in familiar settings, the brain relies on pre-established modules.



	However, this study only considered neuronal connections and did not account for the role of dendritic spines, which also play a part in storage and computation. It is possible that, when faced with familiar stimuli, the nervous system opts for less energy-intensive local changes in dendritic spines to maintain plasticity. This hypothesis awaits further research. Similar research approaches could compare the probability distribution of neuronal density across different brain regions and explore potential differences based on mechanical, genetic, or gene expression factors during development.



	<img alt="" class="aligncenter size-large wp-image-2298" height="912" src="https://admin.next-question.com/wp-content/uploads/2024/08/08-1024x912.webp" width="1024" /> 
	▷ Fig.8. Theoretical and actual probability distributions of neuronal firing rates in rhesus monkeys in response to familiar and unfamiliar stimuli.



	&nbsp;



	A well-known example of the log-normal distribution in the nervous system is the concept of a &quot;neuronal avalanche,&quot; where neural activity experiences brief bursts lasting tens of milliseconds, followed by a few seconds of quiescence. Neuronal avalanches are closely related to the brain&#39;s ability to store, transmit, and compute information, as well as its sparsity and stability. They are among the fundamental constraints to consider when designing neuromorphic computing systems. For example, the &quot;Thousand Brains Theory&quot; proposed for cortical column computation&mdash;whether its model can reproduce the log-normal distribution described here&mdash;could be a criterion for its success.



	Various models based on different assumptions have been developed to explain the mechanisms behind the log-normal distributions observed in the nervous system. Many of these models can fit real data. It is likely that each model captures a particular aspect of the true nature of the nervous system, but the actual pattern of neural activity may result from a combination of these mechanisms.



	From a broader perspective, the log-normal distribution observed in the nervous system reflects its inherent heterogeneity and serves as the foundation for robust learning and energy-efficient computation in the brain. Although terms like &quot;power law,&quot; &quot;scale-free,&quot; and &quot;neuronal avalanche&quot; describe different aspects of these distributions, they fundamentally point to the same complex biological logic. This article provides only an overview of the relevant research; readers interested in these topics can delve further into the references, particularly Reference 1.



	As observed in the stock market simulation game, the brain&#39;s operating patterns bear a notable resemblance to many phenomena in our daily lives, all following certain fundamental mathematical laws. From the micro-level responses of individual neurons to the macro-level dynamics of entire neural networks, the brain utilizes log-normal distributions to optimize information processing, conserve energy, and enhance its adaptability to the external world. This mathematical essence raises the question of whether long-tail distributions and neuronal avalanches, produced by self-organizing processes, can also be observed in artificial neural networks. Do the activation rates in Transformers also follow a log-normal distribution?

[:zh]

[:]

Why Is the Brain Governed by a Log-Normal Distribution?

[:en]Why Is the Brain Governed by a Log-Normal Distribution?[:]

Are You Addicted to GPT?

[:en]Are You Addicted to GPT?[:]

Language and music permeate our daily lives, serving as conduits for communication and emotional expression, and refining our sensibilities. Both involve processing that generates structured representations from basic units. While language and music manifest in different units (such as phonemes, morphemes, notes, and chords), they each combine discrete units into structures with defined hierarchical layers. Studies have found that left-brain regions often associated with music overlap with language areas, including the superior temporal gyrus, inferior frontal gyrus, and inferior parietal lobule.
However, substantive distinctions exist between language and music, including differences in rhythmic structure, pitch usage, the &#8216;meaning&#8217; of representational units, structural building mechanisms, and scope of application. Some argue that music and language are separable cognitive resources. Certain aspects of language, such as speech perception, are distinct from music in the brain. Yet, many other aspects of language processing have not been as extensively explored. Particularly, criteria for measuring the complexity of language and music await further exploration at high spatiotemporal resolution. Given the apparent differences in structural building mechanisms between language and music, specific measures of structural sensitivity may be required to capture domain-specific effects.
To this end, Professor Nitin Tandon and his team have employed intracranial recording methods combined with high spatiotemporal resolution to monitor cortical activity during causal perturbations. They have investigated the spatiotemporal dynamics of music and language structure processing in an epileptic patient. This involved domain-specific complexity analyses and direct cortical stimulation imaging during awake craniotomy. Their findings, published in iScience in July 2023, reveal that music and language, while sharing neural resources, exhibit distinct characteristics when processing domain-specific structures.
&nbsp; 
<img loading="lazy" decoding="async" class="size-large wp-image-1550 alignleft" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-1-1024x579.png" alt="" width="1024" height="579" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/music-1-1024x579.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/02/music-1-300x170.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/music-1-768x434.png 768w, https://admin.next-question.com/wp-content/uploads/2024/02/music-1.png 1063w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
Figure 1: Research Paper Cover. Figure source: iScience official website. 
&nbsp;
<h2>Subjects and Task Paradigms</h2>
The subject of the study is a male professional musician with extensive piano experience dating back to his childhood, who recently began experiencing epileptic seizures leading to transient aphasia. Before the surgery, he completed all the pre-operative tests that were later conducted in the operating room. He excelled in all the language and music tasks during the baseline phase, scoring 172 out of 180 on the Montreal Battery of Evaluation of Amusia (MBEA). Additionally, he engaged in finger tapping exercises on piano keys and practiced freeform playing to familiarize himself with the keyboard setup.
Professor Nitin Tandon and his team collected data from the subject through three behavioral tasks: (1) auditory sentence repetition, where the patient repeated spoken sentences; (2) melody repetition, where he replicated a six-note sequence using one hand on a MIDI keyboard; and (3) auditory naming task, provide one-word answers to definitions of common objects. 
Research Findings
1. ECoG Imaging Characteristics of Music and Language
During the sentence and melody repetition tasks, researchers concurrently recorded the subject&#8217;s specific brain region ECoG activity (Figure 2). In the 100-400 milliseconds post-presentation of each word, 28 lateral electrodes were notably more active than the baseline. These vibrant language electrode sites were distributed across the posterior temporal cortex and inferior frontal cortex. During the melody task, as each note was presented, 11 electrodes within the same time window were significantly active for each tone, mostly clustered around the posterior superior temporal gyrus (pSTG). Seven electrodes were significantly active in both music and language perception, primarily located in the pSTG. In the process of language production, two electrodes in the ventral precentral gyrus were distinctly active.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1551" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-2.png" alt="" width="832" height="306" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/music-2.png 832w, https://admin.next-question.com/wp-content/uploads/2024/02/music-2-300x110.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/music-2-768x282.png 768w" sizes="(max-width: 832px) 100vw, 832px" />
Figure 2: ECoG Activity for Music and Language. Figure source: The paper.
2. Modulation of Broadband High Gamma Activity by the Complexity of Musical and Linguistic Stimuli
For the musical tasks, an electrode in the posterior middle temporal gyrus (pMTG), near the upper boundary of the pSTG, exhibited a significant interaction between complexity and note position. Compared to sequences of lower complexity, high-complexity sequences elicited a greater increase in broadband high gamma activity (Figure 3A). This indicates that the electrode is significantly active in processing both language and music.
When comparing language stimuli based on syntactic complexity, two electrodes on the pSTG demonstrated an increase in BGA (Figure 3B). This suggests that these electrodes are significantly active in both language and music processing. However, the percentage change in BGA during language production was less than that during music production, possibly due to more pronounced differences between musical sequences.
The research also analyzed the low-frequency responses (2-15 Hz and 15-30 Hz) for these same language and music contrasts and found no significant interaction between electrode position and complexity. However, during sentence production, an electrode in the inferior frontal cortex showed a stronger response to syntactic complexity approximately 200-300 milliseconds after the onset of articulation (Figure 3C).
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1552" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-3.png" alt="" width="832" height="794" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/music-3.png 832w, https://admin.next-question.com/wp-content/uploads/2024/02/music-3-300x286.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/music-3-768x733.png 768w" sizes="(max-width: 832px) 100vw, 832px" />
Figure 3: Intraoperative ECoG of Musical and Syntactic Complexity. Figure source: The paper.
&nbsp;
3. Cortical Stimulation Imaging of Language and Music
Beyond ECoG tasks, researchers delved into direct cortical stimulation mapping (CSM) analysis, primarily to observe changes in language and music processing before and after resection in the posterior and middle temporal regions (Figure 4). In the CSM analysis, five sites in the posterior superior temporal gyrus (pSTG) were positive for sentence repetition, leading to disruptions in comprehension. Only two sites in pSTG testing were CSM negative for sentence repetition. In the auditory naming task, three sites in the posterior middle temporal gyrus (pMTG) were CSM positive, with no significant impacts on language task production and comprehension observed.
In the realm of music tasks, CSM primarily characterized music perception or production. Six sites clustered in the pSTG were CSM positive. These six CSM music-positive sites were categorized based on the timing of CSM application during trials. The patient typically introduced &#8220;interference&#8221; while attempting to play or simply failed to select the next note in the melody. Two sites in the ventral temporal cortex were tested for music perception and production, both showing CSM negativity. As depicted in Figure 4, there is a notable overlap between language and music-related sites in the pSTG region. During the CSM language mapping process, the patient&#8217;s subjective reports indicated interference with verbal working memory and, additionally, memory access.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter size-full wp-image-1553" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-4.png" alt="" width="654" height="582" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/music-4.png 654w, https://admin.next-question.com/wp-content/uploads/2024/02/music-4-300x267.png 300w" sizes="(max-width: 654px) 100vw, 654px" />
Figure 4: Intraoperative Stimulation Tests for Music and Language Perception. Figure source: The paper.
&nbsp;
Conclusion
Language and music processing involves effectively combining basic units into structures. It remains to be clarified whether brain regions sensitive to language and music structures are co-located. This study found that stimulating the pSTG disrupted both the perception and production of music, and both the pSTG and pMTG are activated by language and music during language production. Activity in the pMTG is modulated by musical complexity, while activity in the pSTG is modulated by syntactic complexity. This suggests that while music and language share brain resources, they exhibit distinct neural characteristics when processing specific domain structures.
The hallmark of this research is the use of an epileptic subject who is a trained musician, providing a unique model for understanding the neural mechanisms of language and music. Researchers conducted various behavioral tasks and synchronous recordings of intracranial electrophysiological activity on the subject, offering significant insights into the neural underpinnings of these cognitive domains.

[:en]Language and music permeate our daily lives, serving as conduits for communication and emotional expression, and refining our sensibilities. Both involve processing that generates structured representations from basic units. While language and music manifest in different units (such as phonemes, morphemes, notes, and chords), they each combine discrete units into structures with defined hierarchical layers. Studies have found that left-brain regions often associated with music overlap with language areas, including the superior temporal gyrus, inferior frontal gyrus, and inferior parietal lobule.

However, substantive distinctions exist between language and music, including differences in rhythmic structure, pitch usage, the 'meaning' of representational units, structural building mechanisms, and scope of application. Some argue that music and language are separable cognitive resources. Certain aspects of language, such as speech perception, are distinct from music in the brain. Yet, many other aspects of language processing have not been as extensively explored. Particularly, criteria for measuring the complexity of language and music await further exploration at high spatiotemporal resolution. Given the apparent differences in structural building mechanisms between language and music, specific measures of structural sensitivity may be required to capture domain-specific effects.

To this end, Professor Nitin Tandon and his team have employed intracranial recording methods combined with high spatiotemporal resolution to monitor cortical activity during causal perturbations. They have investigated the spatiotemporal dynamics of music and language structure processing in an epileptic patient. This involved domain-specific complexity analyses and direct cortical stimulation imaging during awake craniotomy. Their findings, published in iScience in July 2023, reveal that music and language, while sharing neural resources, exhibit distinct characteristics when processing domain-specific structures.

&nbsp;
<img class="size-large wp-image-1550 alignleft" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-1-1024x579.png" alt="" width="1024" height="579" />
Figure 1: Research Paper Cover. Figure source: iScience official website.
&nbsp;

<h2>Subjects and Task Paradigms</h2>

The subject of the study is a male professional musician with extensive piano experience dating back to his childhood, who recently began experiencing epileptic seizures leading to transient aphasia. Before the surgery, he completed all the pre-operative tests that were later conducted in the operating room. He excelled in all the language and music tasks during the baseline phase, scoring 172 out of 180 on the Montreal Battery of Evaluation of Amusia (MBEA). Additionally, he engaged in finger tapping exercises on piano keys and practiced freeform playing to familiarize himself with the keyboard setup.

Professor Nitin Tandon and his team collected data from the subject through three behavioral tasks: (1) auditory sentence repetition, where the patient repeated spoken sentences; (2) melody repetition, where he replicated a six-note sequence using one hand on a MIDI keyboard; and (3) auditory naming task, provide one-word answers to definitions of common objects.
Research Findings

1. ECoG Imaging Characteristics of Music and Language

During the sentence and melody repetition tasks, researchers concurrently recorded the subject's specific brain region ECoG activity (Figure 2). In the 100-400 milliseconds post-presentation of each word, 28 lateral electrodes were notably more active than the baseline. These vibrant language electrode sites were distributed across the posterior temporal cortex and inferior frontal cortex. During the melody task, as each note was presented, 11 electrodes within the same time window were significantly active for each tone, mostly clustered around the posterior superior temporal gyrus (pSTG). Seven electrodes were significantly active in both music and language perception, primarily located in the pSTG. In the process of language production, two electrodes in the ventral precentral gyrus were distinctly active.

&nbsp;

<img class="aligncenter size-full wp-image-1551" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-2.png" alt="" width="832" height="306" />
Figure 2: ECoG Activity for Music and Language. Figure source: The paper.

2. Modulation of Broadband High Gamma Activity by the Complexity of Musical and Linguistic Stimuli

For the musical tasks, an electrode in the posterior middle temporal gyrus (pMTG), near the upper boundary of the pSTG, exhibited a significant interaction between complexity and note position. Compared to sequences of lower complexity, high-complexity sequences elicited a greater increase in broadband high gamma activity (Figure 3A). This indicates that the electrode is significantly active in processing both language and music.

When comparing language stimuli based on syntactic complexity, two electrodes on the pSTG demonstrated an increase in BGA (Figure 3B). This suggests that these electrodes are significantly active in both language and music processing. However, the percentage change in BGA during language production was less than that during music production, possibly due to more pronounced differences between musical sequences.

The research also analyzed the low-frequency responses (2-15 Hz and 15-30 Hz) for these same language and music contrasts and found no significant interaction between electrode position and complexity. However, during sentence production, an electrode in the inferior frontal cortex showed a stronger response to syntactic complexity approximately 200-300 milliseconds after the onset of articulation (Figure 3C).

&nbsp;

<img class="aligncenter size-full wp-image-1552" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-3.png" alt="" width="832" height="794" />
Figure 3: Intraoperative ECoG of Musical and Syntactic Complexity. Figure source: The paper.
&nbsp;

3. Cortical Stimulation Imaging of Language and Music

Beyond ECoG tasks, researchers delved into direct cortical stimulation mapping (CSM) analysis, primarily to observe changes in language and music processing before and after resection in the posterior and middle temporal regions (Figure 4). In the CSM analysis, five sites in the posterior superior temporal gyrus (pSTG) were positive for sentence repetition, leading to disruptions in comprehension. Only two sites in pSTG testing were CSM negative for sentence repetition. In the auditory naming task, three sites in the posterior middle temporal gyrus (pMTG) were CSM positive, with no significant impacts on language task production and comprehension observed.

In the realm of music tasks, CSM primarily characterized music perception or production. Six sites clustered in the pSTG were CSM positive. These six CSM music-positive sites were categorized based on the timing of CSM application during trials. The patient typically introduced "interference" while attempting to play or simply failed to select the next note in the melody. Two sites in the ventral temporal cortex were tested for music perception and production, both showing CSM negativity. As depicted in Figure 4, there is a notable overlap between language and music-related sites in the pSTG region. During the CSM language mapping process, the patient's subjective reports indicated interference with verbal working memory and, additionally, memory access.

&nbsp;

<img class="aligncenter size-full wp-image-1553" src="https://admin.next-question.com/wp-content/uploads/2024/02/music-4.png" alt="" width="654" height="582" />
Figure 4: Intraoperative Stimulation Tests for Music and Language Perception. Figure source: The paper.
&nbsp;

Conclusion

Language and music processing involves effectively combining basic units into structures. It remains to be clarified whether brain regions sensitive to language and music structures are co-located. This study found that stimulating the pSTG disrupted both the perception and production of music, and both the pSTG and pMTG are activated by language and music during language production. Activity in the pMTG is modulated by musical complexity, while activity in the pSTG is modulated by syntactic complexity. This suggests that while music and language share brain resources, they exhibit distinct neural characteristics when processing specific domain structures.

The hallmark of this research is the use of an epileptic subject who is a trained musician, providing a unique model for understanding the neural mechanisms of language and music. Researchers conducted various behavioral tasks and synchronous recordings of intracranial electrophysiological activity on the subject, offering significant insights into the neural underpinnings of these cognitive domains.[:]

Is Music a Language? High Spatiotemporal Resolution Recordings of the Cerebral Cortex in Search of Answers

[:en]Is Music a Language? High Spatiotemporal Resolution Recordings of the Cerebral Cortex in Search of Answers[:]

[:en]<div style="line-height: 1.5;">The human brain operates with remarkably low energy consumption while performing vast computational processes. Understanding this requires a deep comprehension of the synaptic connections, spatial structures, and static networks of neural circuits, along with the dynamic cognitive processes built upon them. These tasks far exceed the traditional scope of brain region functionality studies. Connectomics research aims to achieve this by precisely mapping how each cell connects with others.
With advancements in scientific research, especially the rapid improvements in computational and data processing capabilities, research methods for the nervous system are also evolving. In this context, technology giants like Google have begun investing in this field, applying advanced technologies to neuroscience research, thereby driving the rapid development of connectomics. Scientists have now mapped detailed connectomes of various organisms' brains, rapidly changing our understanding of brain functioning.
Ten years ago, Google started focusing on AI innovations and collaborating with large datasets, thereby establishing a connectomics team. In reconstructing brain structures, they generated massive amounts of data. To effectively handle this data, the Google team developed the flood-filling networks (FFN) model to replace the traditional manual method of annotating cells in images [1]. These networks can automatically reconstruct neurons through tissue layers. Based on this, Google developed the SegCLR algorithm to automatically identify different parts and cell types within these networks [2]. The Google connectomics team also created TensorStore [3], an open-source C++ and Python software library for storing and managing large multidimensional datasets. This tool has achieved functionality far beyond connectomics and is now widely used at Google and in the broader machine learning community.
In 2018, Google partnered with the Max Planck Institute for Neurobiology in Germany to develop a deep learning-based system that can automatically map brain neurons [1]. They reconstructed the scanned images of one million cubic micrometers of a zebra finch brain.

<img class="aligncenter size-full wp-image-2441" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture1-2.png" alt="" width="999" height="508" />

▷ Illustration of segmenting each neuron in a small part of the zebra finch brain using FFN. Source: google.

In 2019, Google collaborated with the Howard Hughes Medical Institute and the University of Cambridge, using the FFN model and TPU chips to slice a fruit fly brain into thousands of 40-nanometer ultra-thin sections. They used a transmission electron microscope to generate images of each slice, producing over 40 trillion pixels of fruit fly brain images. These 2D images were then aligned to reconstruct a complete 3D image of the fruit fly brain [4]. In 2020, Google released the fruit fly hemibrain connectome, which included images covering 25,000 neurons and revealed their synaptic connections, accounting for about one-third of the fruit fly brain by volume [5].

<img class="aligncenter size-large wp-image-2440" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture2-2-1024x768.png" alt="" width="1024" height="768" />

▷ Fruit fly hemibrain connectome.

Recently, Google collaborated with the Lichtman Laboratory at Harvard University to publish an article in Science titled "A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution."

<img class="aligncenter size-full wp-image-2439" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture3-2.png" alt="" width="951" height="429" />

▷ Shapson-Coe, Alexander, et al. "A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution." Science 384.6696 (2024): eadk4858.

This study released the latest H01 dataset, a 1.4 PB rendered image of a small sample of human brain tissue. The H01 sample was imaged at a resolution of 4nm using serial section electron microscopy and then reconstructed and annotated through automated computational techniques, ultimately revealing the preliminary structure of the human cerebral cortex. The H01 dataset is not only the largest sample of its kind imaged and reconstructed to this extent in any organism, but it is also the first large-scale sample to study the synaptic connectivity of the human cerebral cortex, spanning multiple cell types across all cortical layers.

&nbsp;

<img class="aligncenter size-full wp-image-2438" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture4-2.png" alt="" width="602" height="615" />

▷ Figure 1: The H01 dataset of the human cerebral cortex.

The research team encountered many obstacles in creating this image. One major issue was finding suitable brain tissue samples. Since the human brain begins to degrade rapidly after death, postmortem brain tissue is often unsuitable for research. Therefore, the team chose a tissue sample taken during brain surgery from an epilepsy patient, aiming to help control their seizures. Additionally, identifying the structure of neural circuits at the synaptic level requires higher resolution electron microscopy. Serial EM, with its automation and rapid imaging capabilities, can image cubic millimeter volumes of tissue at nanoscale resolution, making it a powerful tool for reconstructing the fine neural circuits of the human brain.

The outcomes of this work primarily include the following four aspects:

&nbsp;
<h2>1.Reconstruction of Human Brain Samples and Neuron Connectivity</h2>
The human brain sample used in this study was taken from the left anterior temporal lobe of an epilepsy patient during surgery. The sample was about the size of half a grain of rice. After fixation, staining, and embedding, the sample was sectioned using an automatic tape-collecting ultramicrotome (ATUM), producing 5,019 tissue sections with an average thickness of 33.9 nm, totaling 0.170 mm in thickness. These sections were then placed on silicon wafers and imaged under a custom-built 61-beam parallel scanning electron microscope at a resolution of 4x4 nm², rapidly capturing images with a total imaging volume of 1.05 mm³. The raw data size of each section reached up to 350 terabytes, and the overall dataset size reached 1.4 petabytes.

The subsequent computational challenge was to analyze such a massive dataset. The researchers successfully addressed this issue using artificial intelligence algorithms. The team computationally stitched and aligned this data to generate a single 3D volume. Despite the overall high quality of the data, these alignment channels needed to be robust to handle challenges such as imaging artifacts, missing sections, variations in microscope parameters, and physical stretching and compression of the tissue. Once aligned, a multi-scale FFN (flood-filling network) algorithm, utilizing thousands of Google Cloud TPUs, was applied to generate 3D segmentations of each individual cell in the tissue. FFN is the first automated segmentation technique capable of producing sufficiently accurate reconstructions. Additionally, other machine learning pipelines were used to identify and annotate 130 million synapses, dividing each 3D segment into different "subregions" (e.g., axons, dendrites, or cell bodies) and identifying other structures of interest, such as myelin and cilia.

However, the results of automated reconstruction were not perfect. Given the massive amount of neuronal structural data, it was impractical for individual lab members to manually proofread the entire dataset. Therefore, researchers from Princeton University and the Allen Institute developed an online proofreading platform called CAVE (Connectome Annotation and Versioning Engine), as an extension of Neuroglancer. This platform allows users to proofread and update annotated data to improve the dataset as part of their research. Researchers can select specific neurons or their connections and manually refine and annotate their reconstructions. CAVE is already used for smaller datasets, including the fly connectome.

For this project, researchers updated the algorithms to handle such large datasets. The interface permits proofreading, which is a crucial final step in completing the connectome reconstruction. Anyone can apply to become a proofreader, and proofreaders can download data related to the cells they are working on for subsequent analysis. Additionally, VAST (Volume Annotation and Segmentation Tool) is available as a general free software for viewing, segmenting, and annotating large datasets. For diversified analysis, researchers also provided various databases allowing users to query cell and synapse data specifically. To support complex query conditions, they developed a standalone program called CREST (Connectome Reconstruction and Exploration Simple Tool), which identifies and explores cell connections based on many cell characteristics.

<img class="aligncenter size-full wp-image-2437" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture5-2.png" alt="" width="747" height="996" />▷ Source: Viren Jain
<h2>2. Composition of Cells and Synapses in the Human Cerebral Cortex</h2>
Through identification and quantitative analysis of all nucleated cells in the sample, 49,080 neurons and glial cells, along with 8,100 vascular-related cells, were identified. The number of glial cells (32,315) was about twice that of neurons (16,087). These neurons were mainly divided into spiny (65.5%), non-spiny, and non-pyramidal types. The neuronal density in the human cerebral cortex is about 16,000/mm³, which is approximately one-third lower than previously estimated by optical microscopy in the human temporal cortex and nearly ten times lower than the density in the mouse associative cortex.

Based on the size and distribution density of the cell bodies, the researchers divided the sample into six cortical layers and white matter. The distribution of glial cells varied among these layers. Notably, protoplasmic astrocytes were densely distributed from layers 2 to 6. However, in the shallower part of layer 1, they tended to be more mixed, with higher density and smaller size. Fibrous astrocytes in the white matter were more elongated than those occupying the cortex. Two other types of glial cells, microglia and oligodendrocyte precursor cells (OPCs), had nearly uniform density across all layers. Another type of glial cell, oligodendrocytes (n=20,139), exhibited a gradient distribution, with the lowest density in the upper layers and the highest density in the white matter, likely related to their role in myelination. Microglia, OPCs, and oligodendrocytes all showed an affinity for blood vessels. A large number of myelinated axons were radially distributed between the white matter and the superficial cortical layers, horizontally distributed on two orthogonal axes. The reconstructed blood vessels did not show layer-specific distribution.

Synapses are the bridges of the neural network. The human brain has 86 billion neurons, and synapses allow the transmission of electrical signals from one neuron to the next. To identify synaptic sites, the researchers trained a classifier based on the U-Net architecture to label three categories: background, presynaptic, and postsynaptic. They also trained a binary ResNet-50 classifier to categorize each identified synapse as excitatory or inhibitory based on its electron microscopy appearance, postsynaptic structure type, and presynaptic neuron type (if known).

In the reconstructed human brain tissue, approximately 102.5 million (67.1%) synapses were identified as excitatory, and about 50.3 million (32.9%) as inhibitory (Fig.2). Further analysis showed that most of these synapses were located on dendrites (99.4%), with only a small portion on axon initial segments (0.2%) or cell bodies (0.4%). This data is accessible on the Neuroglancer platform. By analyzing the reconstructed synapses, the researchers found that excitatory synapses were mainly distributed in layers 1 and 3, while inhibitory synapses were predominantly found in layer 1. Additionally, the ratio of excitatory to inhibitory inputs received by pyramidal neurons and interneurons varied slightly across layers, reflecting the complex regulatory mechanisms of neural networks.

<img class="aligncenter size-full wp-image-2438" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture4-2.png" alt="" width="602" height="615" />

▷ Figure 2: Distribution of cells and synapses in the human cerebral cortex and white matter.
<h2>3.Morphological Subtypes of Layer 6 Triangular Neurons</h2>
The deeper layers of the cerebral cortex are less understood than the superficial layers due to the complexity and diversity of cell types present. In this study, the researchers focused on spiny neurons in the deep cortical layers, particularly triangular neurons, which are not well understood in primates. In the reconstructed human brain tissue, triangular neurons were predominantly located in layers 5 and 6 (n = 876), accounting for about one-third of the spiny neurons. These cells are characterized by their large basal dendrites extending from the soma in various directions, forming a notable directional distribution, thus referred to as "compass neurons."

The study further distinguished two types of neurons based on the direction of their basal dendrites: one type with large basal dendrites extending anteriorly and another with large basal dendrites extending posteriorly at a mirror-symmetric angle. These two groups of cells formed a mirror-symmetric distribution in the cerebral white matter (Fig.3). Additionally, the direction of the basal dendrites exhibited a bimodal distribution, with most cells extending either significantly forward or markedly backward, while others were more tangentially oriented. The distribution of these basal dendrites corresponded to the functional regions of the cerebral cortex and might be related to the principal axis of the myelinated axons in the subjacent white matter. These neurons tended to cluster with nearest neighbor cells having the same orientation, suggesting that this spatial distribution is not random. The statistical correlation of these mirror-symmetric neurons indicates that such cell clustering might have some underlying, yet unclear, functional significance.

&nbsp;

<img class="aligncenter size-large wp-image-2435" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture7-1-1024x546.png" alt="" width="1024" height="546" />

▷ Figure 3: Two mirror-symmetric subgroups of triangular neurons in the deep layers of the human cerebral cortex.
<h2>4.Multiple Synaptic Connections</h2>
Previous research has shown that axons in the rodent cerebral cortex occasionally form multiple synaptic connections on the same postsynaptic cell. To investigate whether the same phenomenon occurs in the human cerebral cortex, researchers used CREST to systematically identify strong connections in reconstructed human brain tissue. The results indicated that, in the human cerebral cortex, single axons occasionally establish multiple synapses on the same postsynaptic cell. Researchers systematically discovered this phenomenon in both excitatory and inhibitory axons. Although such strong connections are rare, they are extremely potent, with over 50 individual synaptic connections possible between pairs of neurons. Typically, 96.5% of axon-target cell contacts have only one synapse, while 0.092% of axon connections may have four or more synapses.

Furthermore, analysis of 2,743 neurons revealed that 39% of neurons had at least one dendritic input with seven or more synapses, indicating that strong axonal inputs are a common feature of human brain neurons. Interestingly, these multiple synaptic connections are often formed by axons maintaining close contact with dendrites over long distances (Fig.4).

The authors also explored whether the number of synapses established by an axon on target cells exceeded the expectations of a random model. The results showed that the incidence of strong connections observed was significantly higher than the model's predictions (Fig.4). Although the phenomenon of single axons establishing multiple synapses on the same postsynaptic cell is uncommon in the human cerebral cortex, these rare, strong connections might be an important component of complex neuron-to-neuron communication.

<img class="aligncenter size-large wp-image-2434" src="https://admin.next-question.com/wp-content/uploads/2024/09/Picture8-1-1024x871.png" alt="" width="1024" height="871" />

▷ Figure 4: Exceptionally strong synaptic connections in the human cerebral cortex.

Using nanoscale resolution to explore the six-layer structure of the human cortex, researchers revealed details of brain tissue at the cellular to subcellular level, investigating the complex relationships among neurons, synapses, glia, and blood vessels. Additionally, the authors provided several software tools, such as the browser-based Neuroglancer and its derivatives CREST and CAVE, as well as VAST, greatly facilitating data visualization and annotation.

However, the study has certain limitations. Fresh samples from healthy individuals are unlikely to be obtained through this neurosurgical approach. Although the patient's temporal lobe did not show substantial pathological changes under light microscopy, long-term epilepsy or its drug treatment may have subtle effects on cortical tissue connectivity or structure.

Some unusual features were observed in this tissue, including some very large spines, axon varicosities filled with unusual substances, and a few axons forming extensive "axon helices." It is currently unclear whether these rare structures are related to the disease or its treatment or are common features of the human brain. Comparing samples obtained from individuals with different underlying conditions will help us better understand these phenomena. Moreover, since the structure of the human cerebral cortex partly depends on past experiences, there is structural variability among individuals, but the extent of this variability is still unknown.

Considering the greater variability in human experience, behavior, and genetics, and the fact that humans and other vertebrates have many identified neuron categories rather than single identified neuron types, comparing neural circuits among human brains may be more challenging.

</div>[:]

Mapping the Largest Fragment of the Human Brain

[:en]Mapping the Largest Fragment of the Human Brain[:]

In the vanguard of musical inquiry, mainstream theories suggest that music, lacking an explicit evolutionary function, persists in culture due to its implicit social and emotional roles. The importance of such research is evident. Yet, traditional experimental methods in neuroscience often fall short when probing the relationship between music and the brain. Music psychology posits that the controlled environment of the lab contrasts sharply with the dynamic world of live performances and real-life music appreciation. Neuroscientifically, specific brain regions are activated—or more robustly so—only through multimodal input, making the universality of findings under strictly auditory paradigms somewhat unclear.
Recently, the field of music neuroscience has begun pivoting towards more naturalistic and ecologically valid paradigms. Mari Tervaniemi from the University of Helsinki has elucidated this shift through three frameworks: the nature of auditory stimuli and empirical paradigms, the demographics of study participants, and the methodologies and contexts of data collection. This move aims to encourage innovative thinking based on existing naturalistic studies, further advancing the ecological validity of research.
&nbsp; 
<img loading="lazy" decoding="async" class="aligncenter size-large wp-image-1558" src="https://admin.next-question.com/wp-content/uploads/2024/02/listen-1-1024x366.png" alt="" width="1024" height="366" srcset="https://admin.next-question.com/wp-content/uploads/2024/02/listen-1-1024x366.png 1024w, https://admin.next-question.com/wp-content/uploads/2024/02/listen-1-300x107.png 300w, https://admin.next-question.com/wp-content/uploads/2024/02/listen-1-768x274.png 768w, https://admin.next-question.com/wp-content/uploads/2024/02/listen-1.png 1274w" sizes="(max-width: 1024px) 100vw, 1024px" /> 
Figure 1: Research Paper Cover
&nbsp;
<h2>Auditory Stimuli and Empirical Paradigms</h2>
&nbsp; 
In the cerebral concert of auditory neuroscience, sounds used in auditory paradigms are often meticulously crafted, with extraneous variables stringently controlled and stimuli repetitively presented. Researchers encode the sequence of sounds, tagging the auditory stimulus&#8217;s content, and measure the brain&#8217;s response as event-related potentials (ERPs).
However, when longer segments or entire pieces of music serve as auditory stimuli, measuring the brain&#8217;s response to relevant music features requires a different approach. Researchers employ Music Information Retrieval (MIR) tools to quantify acoustic and musical features within the piece and analyze the electroencephalogram (EEG) data as a function of these musical elements. For example, Poikonen and colleagues recorded the EEG signals of musicians, dancers, and laypersons while they listened to a segment from Bizet-Shchedrin&#8217;s Carmen. 
In another condition, participants were presented with an audiovisual segment, including modern dance choreographed to the same piece of music. Using MIR, researchers identified features such as timbral brightness in the music and calculated the corresponding ERPs. They found that dancers had a stronger P50 response to rapid changes in timbral brightness compared to musicians and laypersons. Additionally, when music was paired with dance, the N100 and P200 responses of all participants were both inhibited and accelerated. Thus, employing the MIR toolbox allows for a nuanced neuroscientific study of music over extended durations.
Other studies have utilized shorter musical phrases or beats as stimuli, using event-related potential (ERP) methods to explore the neural underpinnings of musical syntax processing. Like the linguistic P600 response triggered by syntactic violations, an unexpected pitch at the end of a musical phrase can elicit a similar P600 reaction. Another notable ERP is the mismatch negativity (MMN), which measures the neural precision of sound perception and cognition. MMN is triggered by an occasional deviant stimulus interrupting a regular sound sequence. Vuust and colleagues developed an ecologically valid MMN musical paradigm by using an In later MMN paradigms, the notes&#8217; rhythms were looped, with pitches varying randomly.
&nbsp;
<h2>Study Participants</h2>
&nbsp; 
Over the past few decades, the demographics of participants in neuroscience music studies have diversified significantly. Early investigations often compared musicians to laypersons and categorized musicians based on traits like absolute pitch. Later, classifications of musicians became more nuanced, considering factors such as traditional training versus self-taught approaches, as well as musical style and instrument type.
Traditionally, these studies focused on healthy adults. However, the scope has widened to include adults with congenital amusia and individuals with neurological or psychiatric conditions like dystonia, stroke with or without acquired amusia, and depression. More recently, cognitive neuroscience methods have been applied in extensive cross-sectional and longitudinal studies under music medicine and therapy, involving these and other diverse populations.
Beyond adults, children have been recruited for music studies to delve into the cerebral underpinnings of musical learning and understand the transfer effects of musical training on non-musical auditory and cognitive skills, such as language development. Further research has explored music&#8217;s role in treating neurocognitive disorders, including dyslexia and auditory processing disorders.
To understand the impact of early auditory learning, researchers have focused on full-term infants and neonates. Some studies have also involved mothers and their newborns to investigate if prenatal auditory experiences are reflected in the neonates&#8217; EEG as event-related potentials (ERPs) and subcortical frequency following responses.
In summary, data collection in music brain research has expanded to include individuals across the entire lifespan, both neurotypical and atypical. However, most of these studies have focused on Western music, highlighting an urgent need for expanded research into non-Western music. Consequently, the cross-cultural neuroscience of music requires the accumulation of more experience and data from studies involving non-Western musical forms.
&nbsp;
<h2>Data Collection Methods and Context</h2>
&nbsp; 
Technological advancements have extended the application of EEG beyond traditional laboratory settings, allowing for use in environments such as schools and day-care centers. As EEG technology has evolved, it now supports recording and analyzing subjects during broader ranges of movement. This development enables the capture of EEG data from performers playing instruments and listeners moving to music. Subjects can be monitored while playing sequences on a keyboard, with their finger movements, Musical Instrument Digital Interface (MIDI) data, and EEG recorded simultaneously.
Notably, modern concert halls equipped with technology to record physiological signals, including EEG, muscle activity, skin conductance, and heart rate, during live performances, open new avenues for understanding music listening and performance. Facilities like Canada&#8217;s McMaster University&#8217;s LIVELab and the ArtLab at the Max Planck Institute for Empirical Aesthetics in Frankfurt, Germany, can measure EEG and other physiological data in auditoriums and stages. Additionally, laboratories such as the fourMs lab at the University of Oslo in Norway and Casa Paganini at the University of Genoa in Italy are advancing the collection of multimodal data in real-world settings. Finland&#8217;s Aalto University&#8217;s Magicsvi project, which integrates virtual reality into artistic performance and learning, exemplifies the innovative direction of this field. The data from these laboratories, combined with participants&#8217; self-reported experiences like immersion in music and emotional responses, are expected to yield exciting insights into the artistic and scientific facets of music.
It is anticipated that the aforementioned advancements will serve as a catalyst, inspiring further interest and innovation in the study of naturalistic paradigms in music, and providing novel perspectives for exploring the neuroscience of music.

[:en]In the vanguard of musical inquiry, mainstream theories suggest that music, lacking an explicit evolutionary function, persists in culture due to its implicit social and emotional roles. The importance of such research is evident. Yet, traditional experimental methods in neuroscience often fall short when probing the relationship between music and the brain. Music psychology posits that the controlled environment of the lab contrasts sharply with the dynamic world of live performances and real-life music appreciation. Neuroscientifically, specific brain regions are activated—or more robustly so—only through multimodal input, making the universality of findings under strictly auditory paradigms somewhat unclear.

Recently, the field of music neuroscience has begun pivoting towards more naturalistic and ecologically valid paradigms. Mari Tervaniemi from the University of Helsinki has elucidated this shift through three frameworks: the nature of auditory stimuli and empirical paradigms, the demographics of study participants, and the methodologies and contexts of data collection. This move aims to encourage innovative thinking based on existing naturalistic studies, further advancing the ecological validity of research.

&nbsp;
<img class="aligncenter size-large wp-image-1558" src="https://admin.next-question.com/wp-content/uploads/2024/02/listen-1-1024x366.png" alt="" width="1024" height="366" />
Figure 1: Research Paper Cover

&nbsp;
<h2>Auditory Stimuli and Empirical Paradigms</h2>
&nbsp;
In the cerebral concert of auditory neuroscience, sounds used in auditory paradigms are often meticulously crafted, with extraneous variables stringently controlled and stimuli repetitively presented. Researchers encode the sequence of sounds, tagging the auditory stimulus's content, and measure the brain's response as event-related potentials (ERPs).

However, when longer segments or entire pieces of music serve as auditory stimuli, measuring the brain's response to relevant music features requires a different approach. Researchers employ Music Information Retrieval (MIR) tools to quantify acoustic and musical features within the piece and analyze the electroencephalogram (EEG) data as a function of these musical elements. For example, Poikonen and colleagues recorded the EEG signals of musicians, dancers, and laypersons while they listened to a segment from Bizet-Shchedrin's Carmen. 

In another condition, participants were presented with an audiovisual segment, including modern dance choreographed to the same piece of music. Using MIR, researchers identified features such as timbral brightness in the music and calculated the corresponding ERPs. They found that dancers had a stronger P50 response to rapid changes in timbral brightness compared to musicians and laypersons. Additionally, when music was paired with dance, the N100 and P200 responses of all participants were both inhibited and accelerated. Thus, employing the MIR toolbox allows for a nuanced neuroscientific study of music over extended durations.

Other studies have utilized shorter musical phrases or beats as stimuli, using event-related potential (ERP) methods to explore the neural underpinnings of musical syntax processing. Like the linguistic P600 response triggered by syntactic violations, an unexpected pitch at the end of a musical phrase can elicit a similar P600 reaction. Another notable ERP is the mismatch negativity (MMN), which measures the neural precision of sound perception and cognition. MMN is triggered by an occasional deviant stimulus interrupting a regular sound sequence. Vuust and colleagues developed an ecologically valid MMN musical paradigm by using an In later MMN paradigms, the notes' rhythms were looped, with pitches varying randomly.

&nbsp;
<h2>Study Participants</h2>
&nbsp;
Over the past few decades, the demographics of participants in neuroscience music studies have diversified significantly. Early investigations often compared musicians to laypersons and categorized musicians based on traits like absolute pitch. Later, classifications of musicians became more nuanced, considering factors such as traditional training versus self-taught approaches, as well as musical style and instrument type.

Traditionally, these studies focused on healthy adults. However, the scope has widened to include adults with congenital amusia and individuals with neurological or psychiatric conditions like dystonia, stroke with or without acquired amusia, and depression. More recently, cognitive neuroscience methods have been applied in extensive cross-sectional and longitudinal studies under music medicine and therapy, involving these and other diverse populations.

Beyond adults, children have been recruited for music studies to delve into the cerebral underpinnings of musical learning and understand the transfer effects of musical training on non-musical auditory and cognitive skills, such as language development. Further research has explored music's role in treating neurocognitive disorders, including dyslexia and auditory processing disorders.

To understand the impact of early auditory learning, researchers have focused on full-term infants and neonates. Some studies have also involved mothers and their newborns to investigate if prenatal auditory experiences are reflected in the neonates' EEG as event-related potentials (ERPs) and subcortical frequency following responses.

In summary, data collection in music brain research has expanded to include individuals across the entire lifespan, both neurotypical and atypical. However, most of these studies have focused on Western music, highlighting an urgent need for expanded research into non-Western music. Consequently, the cross-cultural neuroscience of music requires the accumulation of more experience and data from studies involving non-Western musical forms.

&nbsp;
<h2>Data Collection Methods and Context</h2>
&nbsp;
Technological advancements have extended the application of EEG beyond traditional laboratory settings, allowing for use in environments such as schools and day-care centers. As EEG technology has evolved, it now supports recording and analyzing subjects during broader ranges of movement. This development enables the capture of EEG data from performers playing instruments and listeners moving to music. Subjects can be monitored while playing sequences on a keyboard, with their finger movements, Musical Instrument Digital Interface (MIDI) data, and EEG recorded simultaneously.

Notably, modern concert halls equipped with technology to record physiological signals, including EEG, muscle activity, skin conductance, and heart rate, during live performances, open new avenues for understanding music listening and performance. Facilities like Canada's McMaster University's LIVELab and the ArtLab at the Max Planck Institute for Empirical Aesthetics in Frankfurt, Germany, can measure EEG and other physiological data in auditoriums and stages. Additionally, laboratories such as the fourMs lab at the University of Oslo in Norway and Casa Paganini at the University of Genoa in Italy are advancing the collection of multimodal data in real-world settings. Finland's Aalto University's Magicsvi project, which integrates virtual reality into artistic performance and learning, exemplifies the innovative direction of this field. The data from these laboratories, combined with participants' self-reported experiences like immersion in music and emotional responses, are expected to yield exciting insights into the artistic and scientific facets of music.

It is anticipated that the aforementioned advancements will serve as a catalyst, inspiring further interest and innovation in the study of naturalistic paradigms in music, and providing novel perspectives for exploring the neuroscience of music.[:]

How might one listen to the brain through music and interpret the neuroscience of music through naturalism?

[:en]How might one listen to the brain through music and interpret the neuroscience of music through naturalism?[:]

What&#8217;s the first thing that comes to mind when you hear about contemplation? Religion, yoga, or mindfulness? As a thinking activity that can reshape the brain and change the ideology, contemplation is at the intersection of cognitive science, human psychology, and cultural psychiatry. Nowadays, contemplation has been gradually separated from religion and increasingly recognized as a healthy training for human both physically and mentally.
&nbsp;
High on the evolutionary tree, human beings have obvious socialization characteristics, and human experience is affected by multiple interactions. On February 25 (Beijing time), the Contemplative Science Summit and Community Gathering, jointly organized by the Tianqiao and Chrissy Chen Institute (TCCI) and the University of California, Davis (UC Davis), was held. During Symposium 2, three scholars from the field of contemplative science explored contemplative practices from the dimensions of eliminating bias in scientific research, neural encoding mechanisms behind language models, and socio-cultural contexts of contemplation.
&nbsp;
<h3>01 Restoring the élan vital of research: Principles and practices of inclusive, anti-racist science</h3>
Kamilah Majied, Ph.D., California State University, Monterey Bay
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1128" src="https://admin.next-question.com/wp-content/uploads/2023/11/50-1-300x170.png" alt="" width="600" height="339" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/50-1-300x170.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/50-1-1024x579.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/50-1-768x434.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/50-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Dr. Kamilah Majied from California State University, Monterey Bay, made a clear statement against racism and immigrant discrimination at the conference. In this symposium, she talked about how to utilize contemplative practices to assess and eliminate bias in scientific research in different social contexts.
&nbsp;
At the beginning, Dr. Majied used a light-hearted contemplative exercise to bring herself closer to the audience and create a relaxed and joyful atmosphere for the symposium. Dr. Majied argued that it is beneficial to pause when we are in an emotionally and cognitively difficult moment. Such pauses will create surprises and serves as a buffer for unconscious biases.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1129" src="https://admin.next-question.com/wp-content/uploads/2023/11/51-1-300x169.png" alt="" width="600" height="337" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/51-1-300x169.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/51-1-1024x576.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/51-1-768x432.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/51-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Dr. Majied described her own passion for resisting white supremacy culture and her expectations that people will explore science and the world fully and freely. She prescribed her own antidotes in terms of a culture of appreciation, a framework for learning, and positive feedback as well as criticism. She argues that haste breeds discrimination, while momentary cease leaves space for sanity. White supremacy in science and education should be ignored rather than being a hindrance to the road forward. Language, as a vehicle of culture, allows for authentic and holistic communication and should be an effective practice against racism.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1130" src="https://admin.next-question.com/wp-content/uploads/2023/11/52-1-300x170.png" alt="" width="600" height="339" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/52-1-300x170.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/52-1-1024x579.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/52-1-768x434.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/52-1.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
&#8220;Let thoughts flow freely and words turn into poetry to nourish our hearts and minds. It delights the wise to create a better world with all.&#8221; Towards the end of his talk, Dr. Majied engaged the audience in a fill-in-the-blank interaction, where everyone was free to articulate their thoughts. While prejudice and discrimination is a heavy topic, Dr. Majied took a transcendental approach to address existing problems and offered viable solutions in conjunction with the contemplative science.
&nbsp;
<h3>02 Can deep language models serve as a cognitive model for natural language processing in the human brain?</h3>
Uri Hasson, Ph.D., Princeton University
&nbsp;
Dr. Uri Hasson from Princeton University talked about the neural processes of natural language processing and shared the computational rules shared by neural encoding and deep language models.
&nbsp;
Dr. Hasson pointed out that research subjects in the lab usually cannot be used as references for real-life scenarios, and the deep language model he initially constructed has been proven to fail in multi-scenario validation. He has conducted a series of explorations ever since to construct a model in recent years that is close to language processing in the natural state of the human brain.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1131" src="https://admin.next-question.com/wp-content/uploads/2023/11/53-300x169.png" alt="" width="600" height="339" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/53-300x169.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/53-1024x578.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/53-768x434.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/53.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
However, current deep language models still have shortcomings, such as the lack of advanced planning and the limitation that the language learning process is carried out by observing word-to-word connections. These flaws indicate that such models are still greatly different from the human brain.
&nbsp;
This topic is naturally associated with ChatGPT, the immensely popular large language model launched recently, which leads to an interesting question: how many computational laws are shared between ChatGPT and the human brain in terms of language processing?
&nbsp;
<h3>03 Contemplative practice in context: A cultural-ecosocial approach</h3>
Laurence Kirmayer, M.D., McGill University
&nbsp;
Dr. Laurence Kirmayer from McGill University, USA, discussed the cultural shaping mechanisms involved in contemplative practices and reflected on contemplative science under social, cultural, and political contexts.
&nbsp;
Dr. Kirmayer argued that the subject of contemplative science consists of four parts: trainings, processes, goals, and results, of which contemplative trainings can take many forms depending on cultural diversities.
&nbsp;
The social sciences provide contemplative trainings based on specific social contexts, and contemplation in turn can help individuals recognize the dynamics of social systems. The continuation of human society is not only dependent on our brains, but also our interactions with the environment.
&nbsp;
<img loading="lazy" decoding="async" class="aligncenter wp-image-1132" src="https://admin.next-question.com/wp-content/uploads/2023/11/54-300x164.png" alt="" width="600" height="327" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/54-300x164.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/54-1024x558.png 1024w, https://admin.next-question.com/wp-content/uploads/2023/11/54-768x419.png 768w, https://admin.next-question.com/wp-content/uploads/2023/11/54.png 1080w" sizes="(max-width: 600px) 100vw, 600px" />
&nbsp;
Contemplative thinking changes dynamically from context to context. And in the encoding process of daily attention, the interactions of self-consciousness, consciousness of others, illusions, and subconsciousness are maneuvering our mental worlds. Neuroscientists are constantly interpreting the biological process behind it, but it is indisputable that its carrier is the very continuation of the physical body and culture.
&nbsp;
Contemplation does not exist in isolation but intertwine with specific historical and social contexts to reveal its uniqueness. Dr. Kirmayer demonstrated the connotations of contemplation in specific contexts from multiple dimensions. Finally, he advocated that the study of the contemplative science should be an integral part of a cultural system, with an eye on changes caused by political shifts.
&nbsp;
Though the three speakers presented around the same topic of contemplation, each took on different perspectives including bias in science and culture and specific cultural contexts, which unveiled the true beauty of contemplation. The neural mechanism of language encoding is a unique perspective to rationally analyze the biological basis of contemplation. We aspire to demystify contemplation and usher it into the scientific realm in the form of mindfulness for the benefit of all mankind.

[:en]What's the first thing that comes to mind when you hear about contemplation? Religion, yoga, or mindfulness? As a thinking activity that can reshape the brain and change the ideology, contemplation is at the intersection of cognitive science, human psychology, and cultural psychiatry. Nowadays, contemplation has been gradually separated from religion and increasingly recognized as a healthy training for human both physically and mentally.

&nbsp;

High on the evolutionary tree, human beings have obvious socialization characteristics, and human experience is affected by multiple interactions. On February 25 (Beijing time), the Contemplative Science Summit and Community Gathering, jointly organized by the Tianqiao and Chrissy Chen Institute (TCCI) and the University of California, Davis (UC Davis), was held. During Symposium 2, three scholars from the field of contemplative science explored contemplative practices from the dimensions of eliminating bias in scientific research, neural encoding mechanisms behind language models, and socio-cultural contexts of contemplation.

&nbsp;
<h3>01 Restoring the élan vital of research: Principles and practices of inclusive, anti-racist science</h3>
Kamilah Majied, Ph.D., California State University, Monterey Bay

&nbsp;

<img class="aligncenter wp-image-1128" src="https://admin.next-question.com/wp-content/uploads/2023/11/50-1-300x170.png" alt="" width="600" height="339" />

&nbsp;

Dr. Kamilah Majied from California State University, Monterey Bay, made a clear statement against racism and immigrant discrimination at the conference. In this symposium, she talked about how to utilize contemplative practices to assess and eliminate bias in scientific research in different social contexts.

&nbsp;

At the beginning, Dr. Majied used a light-hearted contemplative exercise to bring herself closer to the audience and create a relaxed and joyful atmosphere for the symposium. Dr. Majied argued that it is beneficial to pause when we are in an emotionally and cognitively difficult moment. Such pauses will create surprises and serves as a buffer for unconscious biases.

&nbsp;

<img class="aligncenter wp-image-1129" src="https://admin.next-question.com/wp-content/uploads/2023/11/51-1-300x169.png" alt="" width="600" height="337" />

&nbsp;

Dr. Majied described her own passion for resisting white supremacy culture and her expectations that people will explore science and the world fully and freely. She prescribed her own antidotes in terms of a culture of appreciation, a framework for learning, and positive feedback as well as criticism. She argues that haste breeds discrimination, while momentary cease leaves space for sanity. White supremacy in science and education should be ignored rather than being a hindrance to the road forward. Language, as a vehicle of culture, allows for authentic and holistic communication and should be an effective practice against racism.

&nbsp;

<img class="aligncenter wp-image-1130" src="https://admin.next-question.com/wp-content/uploads/2023/11/52-1-300x170.png" alt="" width="600" height="339" />

&nbsp;

"Let thoughts flow freely and words turn into poetry to nourish our hearts and minds. It delights the wise to create a better world with all." Towards the end of his talk, Dr. Majied engaged the audience in a fill-in-the-blank interaction, where everyone was free to articulate their thoughts. While prejudice and discrimination is a heavy topic, Dr. Majied took a transcendental approach to address existing problems and offered viable solutions in conjunction with the contemplative science.

&nbsp;
<h3>02 Can deep language models serve as a cognitive model for natural language processing in the human brain?</h3>
Uri Hasson, Ph.D., Princeton University

&nbsp;

Dr. Uri Hasson from Princeton University talked about the neural processes of natural language processing and shared the computational rules shared by neural encoding and deep language models.

&nbsp;

Dr. Hasson pointed out that research subjects in the lab usually cannot be used as references for real-life scenarios, and the deep language model he initially constructed has been proven to fail in multi-scenario validation. He has conducted a series of explorations ever since to construct a model in recent years that is close to language processing in the natural state of the human brain.

&nbsp;

<img class="aligncenter wp-image-1131" src="https://admin.next-question.com/wp-content/uploads/2023/11/53-300x169.png" alt="" width="600" height="339" />

&nbsp;

However, current deep language models still have shortcomings, such as the lack of advanced planning and the limitation that the language learning process is carried out by observing word-to-word connections. These flaws indicate that such models are still greatly different from the human brain.

&nbsp;

This topic is naturally associated with ChatGPT, the immensely popular large language model launched recently, which leads to an interesting question: how many computational laws are shared between ChatGPT and the human brain in terms of language processing?

&nbsp;
<h3>03 Contemplative practice in context: A cultural-ecosocial approach</h3>
Laurence Kirmayer, M.D., McGill University

&nbsp;

Dr. Laurence Kirmayer from McGill University, USA, discussed the cultural shaping mechanisms involved in contemplative practices and reflected on contemplative science under social, cultural, and political contexts.

&nbsp;

Dr. Kirmayer argued that the subject of contemplative science consists of four parts: trainings, processes, goals, and results, of which contemplative trainings can take many forms depending on cultural diversities.

&nbsp;

The social sciences provide contemplative trainings based on specific social contexts, and contemplation in turn can help individuals recognize the dynamics of social systems. The continuation of human society is not only dependent on our brains, but also our interactions with the environment.

&nbsp;

<img class="aligncenter wp-image-1132" src="https://admin.next-question.com/wp-content/uploads/2023/11/54-300x164.png" alt="" width="600" height="327" />

&nbsp;

Contemplative thinking changes dynamically from context to context. And in the encoding process of daily attention, the interactions of self-consciousness, consciousness of others, illusions, and subconsciousness are maneuvering our mental worlds. Neuroscientists are constantly interpreting the biological process behind it, but it is indisputable that its carrier is the very continuation of the physical body and culture.

&nbsp;

Contemplation does not exist in isolation but intertwine with specific historical and social contexts to reveal its uniqueness. Dr. Kirmayer demonstrated the connotations of contemplation in specific contexts from multiple dimensions. Finally, he advocated that the study of the contemplative science should be an integral part of a cultural system, with an eye on changes caused by political shifts.

&nbsp;

Though the three speakers presented around the same topic of contemplation, each took on different perspectives including bias in science and culture and specific cultural contexts, which unveiled the true beauty of contemplation. The neural mechanism of language encoding is a unique perspective to rationally analyze the biological basis of contemplation. We aspire to demystify contemplation and usher it into the scientific realm in the form of mindfulness for the benefit of all mankind.[:zh]听到冥想，你首先想到的是什么呢？宗教、瑜伽、物我两忘？冥想是一种可以重塑大脑、改变意识形态的思维活动，交汇于认知科学、人类心理学与文化精神病学。当下，冥想逐渐从宗教中分离出来，被越来越多人所认可，成为有益于人类身心健康的训练。

&nbsp;

人类作为进化树上的高级存在，具有明显的社会化特征，人类体验受到多种交互作用的影响。北京时间2月25日，天桥脑科学研究院（TCCI）与美国加州大学戴维斯分校（UC Davis）联合举办的冥想科学会议顺利召开。在“专题研讨会二”中，来自冥想科学领域的3位学者从消除科学研究中的偏见、语言模型的神经编码机制以及冥想的社会文化背景等多个维度对冥想实践进行探讨。

&nbsp;
<h3>01</h3>
<h3>重塑研究的生命力：反种族歧视的包容性科学的原则与实践</h3>
&nbsp;

[caption id="attachment_1128" align="aligncenter" width="600"]<img class="wp-image-1128" src="https://admin.next-question.com/wp-content/uploads/2023/11/50-1-300x170.png" alt="" width="600" height="339" /> ▷图注：演讲主题[/caption]

&nbsp;

来自美国加州州立大学蒙特雷湾分校的Kamilah Majied博士在会上明确表示，反对种族主义与移民歧视。在此次研讨会上，她讲述了如何利用冥想实践，评估和消除不同社会背景下科学研究中存在的偏见。

&nbsp;

一开始，Majied博士用一段轻松愉快的冥想训练搭建了与在场听众互动的桥梁，营造了轻松快乐的会议氛围。Majied博士认为，当我们的情感和认知处于艰难时刻时，不妨停下来。这种停歇会创造惊喜，为无意识偏见提供缓冲空间。

&nbsp;

[caption id="attachment_1129" align="aligncenter" width="600"]<img class="wp-image-1129" src="https://admin.next-question.com/wp-content/uploads/2023/11/51-1-300x169.png" alt="" width="600" height="337" /> ▷图注：讲者对改善偏见提供的观点[/caption]

&nbsp;

Majied博士介绍了自己抵抗白人优先文化的热忱，期望人们可以全面、自由地探索科学与世界。她从感恩文化、学习框架、批评与积极反馈等方面提出了解决之道。她认为，急切给歧视提供土壤，片刻的停歇创造理智，科学与教育中的白人优先应该被忽略而不应成为前行的负担。语言作为文化的载体，可以真实、全面地进行沟通，应该成为反对种族主义的有效实践。

&nbsp;

[caption id="attachment_1130" align="aligncenter" width="600"]<img class="wp-image-1130" src="https://admin.next-question.com/wp-content/uploads/2023/11/52-1-300x170.png" alt="" width="600" height="339" /> ▷图注：作者用以阐明观点的诗句[/caption]

&nbsp;

“让思想在心头自由地流淌，让文字化作诗歌滋润我们的心扉，能让智者愉悦的是与众人一起创造更美好的世界。”在演讲的尾声，Majied博士与在场听众进行了填词互动，大家可以自由阐述自己的想法。虽然偏见与歧视是一个沉重的话题，但Majied博士以超然物外的态度陈述了存在的问题并结合冥想科学提供了可行的解决方案。

&nbsp;
<h3>02</h3>
<h3>深度语言模型能否作为人脑自然语言处理的认知模型？</h3>
&nbsp;

来自美国普林斯顿大学的Uri Hasson博士讲述了自然语言处理的神经过程，并分享了神经编码与深度语言模型共享的计算规则。

&nbsp;

Hasson博士指出，实验室里的研究对象往往不能作为现实生活场景的参考，他最初构建的深度语言模型在多场景验证中失效了，此后经历了一系列探索，近年来构建的模型接近人脑自然状态下的语言加工过程。

&nbsp;

[caption id="attachment_1131" align="aligncenter" width="600"]<img class="wp-image-1131" src="https://admin.next-question.com/wp-content/uploads/2023/11/53-300x169.png" alt="" width="600" height="339" /> ▷图注：深度语言模型与人脑的区别[/caption]

&nbsp;

但当前的深度语言模型仍存在不足，如缺乏提前规划；语言学习过程是通过观察单词与单词之间的联系进行的，具有局限性；与人脑仍有较多差别。

&nbsp;

这不禁令人联想到最近广受关注的ChatGPT，并引申出一个有趣的问题：ChatGPT与人脑在语言加工过程中又有多少共享的计算法则呢？

&nbsp;
<h3>03</h3>
<h3>具体情景下的冥想实践：基于文化与生态社会的实践方式</h3>
&nbsp;

来自美国麦吉尔大学的Laurence Kirmayer博士讲述了冥想实践涉及的文化塑造机制，从社会、文化、政治背景的层面对冥想科学进行思考。

&nbsp;

Kirmayer博士认为冥想科学的主题由训练、过程、目标、结果四部分组成，其中冥想训练会因文化的多元而呈现多种形式。

&nbsp;

社会科学提供基于特定社会环境的冥想训练，冥想反过来可以帮助个体认识到社会系统的动态变化。人类社会的延续不仅依赖于我们的大脑，同时也包含了与环境的交融。

&nbsp;

[caption id="attachment_1132" align="aligncenter" width="600"]<img class="wp-image-1132" src="https://admin.next-question.com/wp-content/uploads/2023/11/54-300x164.png" alt="" width="600" height="327" /> ▷图注：意识、经验、社会因素与神经生物学过程的相互作用[/caption]

&nbsp;

不同情境下，冥想的思维水平是动态变化的。而在日常注意力的编码过程中，自我意识、他人意识、幻想、潜意识四者相互作用，舞动着人们的精神世界，神经科学家们正不断诠释背后的生物学过程，但毋庸置疑，其载体正是延续的肉体与文化。

&nbsp;

冥想不是孤立存在的，它与特定的历史、社会情境交融，展现其独特的一面。Kirmayer博士从多个维度展示了特定情境下的冥想内涵。最后，他倡议将冥想科学作为文化系统的一部分进行研究，同时应注意政治变动带来的变化。

&nbsp;

本专题的三场演讲同为冥想主题，却分别带着我们透过科学文化的偏见、特定的文化情境，感受冥想的真正魅力。语言编码的神经机制则从独特的视角对冥想的生物学基础进行理性剖析。我们期望冥想能揭开神秘的面纱，以正念的形式走入科学研究，造福全人类。

&nbsp;[:]

A joint UC Davis-TCCI Contemplative Science Summit-Symposium 2: Traditional labs may limit contemplative research?

[:en]A joint UC Davis-TCCI Contemplative Science Summit-Symposium 2: Traditional labs may limit contemplative research?[:zh]传统实验室或许是冥想研究的樊笼？冥想科学峰会·专题二[:]

Is transcending boundaries the sole path for embodied intelligence?

[:en]Is transcending boundaries the sole path for embodied intelligence?[:zh]突破边界线，或是具身智能的唯一出路？[:]

TCCI Hosts First Interdisciplinary Meeting to Explore Integration between AI and Mental Health Research
&nbsp;
In a German test in April this year, GPT-3.5 model showed that it scored high on commonly used questionnaires for testing anxiety and demonstrated specific decision-making and biased behaviors when prompted in an anxiety-inducing situation. This raised concerns about whether machines can suffer from emotional disorders such as anxiety, just as humans do.
&nbsp;
On May 25, the Tianqiao and Chrissy Chen Institute (TCCI) hosted an online discussion called &#8220;AI and the Brain.” Featuring Jianyin Qiu, Physician and Director of the Center for Psychological Counselling and Treatment at the Shanghai Mental Health Center, and Guang Chen, Associate Professor at the School of Artificial Intelligence at Beijing University of Posts and Telecommunications, the event attracted around 70,000 live audience members.
&nbsp;
As a senior psychiatrist and psychotherapist, Jianyin Qiu is excited to see that new AI technology has taken its performance in emotional cognition to the next level. She’s very open-minded about and even looking forward to making AI suffer from anxiety after professional training, so as to help doctors better understand and analyze the symptoms and factors associated with the condition. This will open up new ideas for finding more and better approaches to prevent and cure the disorder.
&nbsp;
On the other hand, Guang Chen, who has already gained much attention by sharing AI-related research and progress on social media for years, believes that such models are based on a large amount of training data to learn and generate text. The models trained in a way similar to fill-in-the-blank and solitaire do not currently possess authentic subjective consciousness, emotions or will. Relying on reinforcement learning from human feedback (RLHF), AI technology can indeed continuously adapt its answers to meet the expectations of human prompters in accordance with their responses and guidance. However, he argued that currently, such imitations neither went beyond text learning and generation, nor did they form a stable and consistent &#8216;personality&#8217;.
&nbsp;
According to Jianyin Qiu, pathological anxiety is a persistent, unspecified state of nervousness or groundless foreboding of apocalypse, jeopardy, or imminent catastrophe, accompanied by significant dysautonomia and motor restlessness. It’ll often result in subjective distress or impaired social functioning. Statistics from 2020 released by the National Health Commission of China show that the prevalence rate of anxiety disorders in China has reached around 5%.
 
In recent years, doctors have been proactively experimenting with the application of AI technology in the diagnosis and treatment of mental disorders, including anxiety. As of now, AI has already played a positive role in areas such as suicide prevention. As the scope of human-machine interaction expands from text-based dialogues to integrate facial expressions, body posture, audio, heart rate, body temperature, and even brain-computer interfaces in the future, AI will help further reveal the physiological mechanisms behind human emotions. Technological advancement in the AI field may be translated into accessible products such as a virtual psychotherapist that can assist mental health treatment in residential communities or remote areas with scarce medical resources.
&nbsp;
From the perspective of computer science, Guang Chen observed that the ChatGPT boom has encouraged the use of AI in various industries. If ‘fine-tuned’ by data from the mental health domain, large language models will optimize the AI simulations of human emotions and play the role of mentally disturbed patients for medical research. In addition, it is expected to identify emotional disorders more accurately and efficiently, which can aid doctors in diagnosis and treatment. The scale of relevant data is crucial to the application of AI technology in the field of mental health.
&nbsp;
Both Jianyin Qiu and Guang Chen believe that traditionally, doctors&#8217; understanding of mental disorders have relied heavily on subjective observation. However, the emerging field of computational psychiatry has been driving the application of tools such as data analytics, machine learning and artificial intelligence in the diagnostic typing, risk warning and prognosis prediction of mental disorders, as well as in formulating clinical diagnosis, treatment, and prevention strategies. Large language models, represented by ChatGPT, will bring even greater breakthroughs in computational psychiatry and further decode psychosis and the human brain for the benefit of humanity.
&nbsp;
<div id="attachment_1097" style="width: 610px" class="wp-caption aligncenter"><img loading="lazy" decoding="async" aria-describedby="caption-attachment-1097" class="wp-image-1097" src="https://admin.next-question.com/wp-content/uploads/2023/11/21-2-300x300.png" alt="" width="600" height="600" srcset="https://admin.next-question.com/wp-content/uploads/2023/11/21-2-300x300.png 300w, https://admin.next-question.com/wp-content/uploads/2023/11/21-2-150x150.png 150w, https://admin.next-question.com/wp-content/uploads/2023/11/21-2.png 400w" sizes="(max-width: 600px) 100vw, 600px" />Scan the QR code to watch the replay</div>

[:en]TCCI Hosts First Interdisciplinary Meeting to Explore Integration between AI and Mental Health Research

&nbsp;

In a German test in April this year, GPT-3.5 model showed that it scored high on commonly used questionnaires for testing anxiety and demonstrated specific decision-making and biased behaviors when prompted in an anxiety-inducing situation. This raised concerns about whether machines can suffer from emotional disorders such as anxiety, just as humans do.

&nbsp;

On May 25, the Tianqiao and Chrissy Chen Institute (TCCI) hosted an online discussion called "AI and the Brain.” Featuring Jianyin Qiu, Physician and Director of the Center for Psychological Counselling and Treatment at the Shanghai Mental Health Center, and Guang Chen, Associate Professor at the School of Artificial Intelligence at Beijing University of Posts and Telecommunications, the event attracted around 70,000 live audience members.

&nbsp;

As a senior psychiatrist and psychotherapist, Jianyin Qiu is excited to see that new AI technology has taken its performance in emotional cognition to the next level. She’s very open-minded about and even looking forward to making AI suffer from anxiety after professional training, so as to help doctors better understand and analyze the symptoms and factors associated with the condition. This will open up new ideas for finding more and better approaches to prevent and cure the disorder.

&nbsp;

On the other hand, Guang Chen, who has already gained much attention by sharing AI-related research and progress on social media for years, believes that such models are based on a large amount of training data to learn and generate text. The models trained in a way similar to fill-in-the-blank and solitaire do not currently possess authentic subjective consciousness, emotions or will. Relying on reinforcement learning from human feedback (RLHF), AI technology can indeed continuously adapt its answers to meet the expectations of human prompters in accordance with their responses and guidance. However, he argued that currently, such imitations neither went beyond text learning and generation, nor did they form a stable and consistent 'personality'.

&nbsp;

According to Jianyin Qiu, pathological anxiety is a persistent, unspecified state of nervousness or groundless foreboding of apocalypse, jeopardy, or imminent catastrophe, accompanied by significant dysautonomia and motor restlessness. It’ll often result in subjective distress or impaired social functioning. Statistics from 2020 released by the National Health Commission of China show that the prevalence rate of anxiety disorders in China has reached around 5%.

 

In recent years, doctors have been proactively experimenting with the application of AI technology in the diagnosis and treatment of mental disorders, including anxiety. As of now, AI has already played a positive role in areas such as suicide prevention. As the scope of human-machine interaction expands from text-based dialogues to integrate facial expressions, body posture, audio, heart rate, body temperature, and even brain-computer interfaces in the future, AI will help further reveal the physiological mechanisms behind human emotions. Technological advancement in the AI field may be translated into accessible products such as a virtual psychotherapist that can assist mental health treatment in residential communities or remote areas with scarce medical resources.

&nbsp;

From the perspective of computer science, Guang Chen observed that the ChatGPT boom has encouraged the use of AI in various industries. If ‘fine-tuned’ by data from the mental health domain, large language models will optimize the AI simulations of human emotions and play the role of mentally disturbed patients for medical research. In addition, it is expected to identify emotional disorders more accurately and efficiently, which can aid doctors in diagnosis and treatment. The scale of relevant data is crucial to the application of AI technology in the field of mental health.

&nbsp;

Both Jianyin Qiu and Guang Chen believe that traditionally, doctors' understanding of mental disorders have relied heavily on subjective observation. However, the emerging field of computational psychiatry has been driving the application of tools such as data analytics, machine learning and artificial intelligence in the diagnostic typing, risk warning and prognosis prediction of mental disorders, as well as in formulating clinical diagnosis, treatment, and prevention strategies. Large language models, represented by ChatGPT, will bring even greater breakthroughs in computational psychiatry and further decode psychosis and the human brain for the benefit of humanity.

&nbsp;

[caption id="attachment_1097" align="aligncenter" width="600"]<img class="wp-image-1097" src="https://admin.next-question.com/wp-content/uploads/2023/11/21-2-300x300.png" alt="" width="600" height="600" /> Scan the QR code to watch the replay[/caption][:zh]ChatGPT在“智慧涌现”的同时，会不会涌现出情绪？今年4月一项德国研究显示，GPT-3.5在人类常用的焦虑症测试问卷中获得了较高的分数，且被提示诱导进入焦虑情景后，如研究人员所预期的那样产生了特定的决策判断和偏见行为。这是不是意味着，AI有可能经过训练患上焦虑症？

&nbsp;

5月25日，在天桥脑科学研究院（Tianqiao &amp; Chrissy Chen Institute, TCCI）举办的首期“AI问脑”对话上，上海市精神卫生中心心理咨询与治疗中心主任仇剑崟主任医师与北京邮电大学人工智能学院副教授陈光就此话题在线交流互动，吸引了近7万人次直播观众。

&nbsp;

作为资深精神科医师、心理治疗师，仇剑崟很兴奋地看到，AI这次在情绪认知方面的表现上了一个新台阶。她坦言，自己很期待AI经过专业训练之后能够患上焦虑症，从而帮助医生更好地了解焦虑症的表现，分析背后的相关因素，为找出更多更好的预防、治愈方法打开新思路。

&nbsp;

陈光在微博上的ID是@爱可可-爱生活，多年来致力于AI相关领域的研究和分享，获得了广泛关注。他认为，这类模型基于大量的训练数据进行学习和生成文本。用类似填空、接龙的方式训练得到的模型，目前不具备真正的主观意识，没有真正的主观意识、情绪或意愿。而基于人工反馈强化学习（RLHF），AI的确可根据人类对其输出结果的判定和引导，不断调整，逐渐接近人类的倾向，但他认为，目前看来，这种模仿没有超出文本学习和生成的范围，并不会形成稳定、一致的“性格”。

&nbsp;

仇剑崟介绍，病理性焦虑是指持续、无具体原因地感到紧张不安，或无现实依据地预感到灾难、威胁或大祸临头感，伴有明显的自主神经功能紊乱及运动性不安，常常伴随主观痛苦感或社会功能受损。国家卫健委2020年统计数据显示，中国焦虑症患病率在5%左右。

&nbsp;

近年来，医生开始积极尝试将AI技术应用于包括焦虑症在内的精神障碍临床诊疗。目前，AI已经在自杀预防等领域起到积极的辅助作用。随着人机交互的范围从文本拓展到面部表情、身体姿态、语音、心率、体温，甚至是未来的脑机接口，AI将进一步帮助揭示人类情绪背后的生理学机制，成为心理治疗机器人，在社区或医疗资源匮乏的边远地区助力精神卫生诊疗。

&nbsp;

陈光从计算机科学的角度观察到，ChatGPT热潮促进AI进入各行各业广泛应用，如果结合精神卫生领域的专业数据对大模型进行“微调”，一方面将帮助AI更好地模拟人类的情绪反应，模拟出一批供医生研究的“精神病人”，另一方面也有望更准确高效地识别出人类的情绪障碍，起到辅助诊疗作用。相关数据的规模是AI技术能否深入精神卫生领域的关键。

&nbsp;

仇剑崟和陈光都认为，传统上，医生对精神障碍的认知依赖主观观察。计算精神病学的兴起，推动数据分析、机器学习和人工智能等工具应用于精神障碍的分型诊断、风险预警、预后预测，协助临床诊疗和防治策略制定。而以ChatGPT为代表的大模型AI，将给计算精神病学带来更大的突破，帮助人类进一步打开精神疾病和大脑这两个“黑盒”。

&nbsp;
<h3>结语</h3>
&nbsp;

TCCI致力于支持全球范围内的脑科学交流，仅2022年就主办、合办、支持了近200场会议，遍及北美、亚洲、欧洲和大洋洲。“AI问脑”是一档新推出的轻型学术交流会议系列，以AI科学家和脑科学家交叉对话为主，探索AI技术与脑科学的双向创新，并对AI自主意识、人类意识上云等公众热议话题进行开放性讨论。

&nbsp;

TCCI由盛大集团创始人，中国网络游戏、网络文学行业开创者陈天桥、雒芊芊夫妇出资10亿美元创建，聚焦AI＋脑科学，支持、推进全球范围内脑科学研究，造福人类。TCCI一期投入5亿元人民币支持中国脑科学研究，与上海周良辅医学发展基金会合作成立上海陈天桥脑健康研究所，与华山医院、上海市精神卫生中心等建立战略合作，设立了应用神经技术前沿实验室、人工智能与精神健康前沿实验室。在国际上，TCCI与加州理工学院合作成立TCCI加州理工研究院。[:]

Will AI Have Anxiety?

[:en]Will AI Have Anxiety? [:zh]AI会得焦虑症吗？天桥脑科学研究院举办首期“AI问脑”，跨界探讨AI与精神卫生研究融合创新[:]

Acupuncture: The Revival of Ancient Wisdom in Modern Neuroscience

[:en]Acupuncture: The Revival of Ancient Wisdom in Modern Neuroscience[:]

Recently, scientists at UC Berkeley have succeeded in reconstructing the music the brain hears from brain waves. The results came from UC Berkeley&#8217;s Helen Wills Neuroscience Laboratory, and Professor Gerwin Schalk, Director of the Chen Frontier Lab for Applied Neurotechnology, was one of the researchers.
&nbsp;
A total of 29 patients with drug-resistant epilepsy who had undergone intracranial electrode implantation to monitor seizures were included in this study, with a total of 2,668 electrodes implanted. The researchers asked the patients to listen to a certain piece of music and thereafter used artificial intelligence to decode and reconstruct the information recorded by the electrodes.
&nbsp;
<h3>Which brain regions have registered music?</h3>
&nbsp;
To identify electrodes encoding acoustical information about the song, the researchers used artificial intelligence to fit spectrotemporal receptive fields for all 2,379 artifact-free electrodes in the dataset, assessing how well the high-frequency activity (HFA) recorded at these sites could be linearly predicted from the song’s auditory spectrogram: the better the prediction, the more the electrodes at that location are associated with music recording.
&nbsp;
<h3>How much information is required to reconstruct musi from the brain?</h3>
&nbsp;
The researchers found that 80% of the best prediction accuracy was obtained with 43 electrodes; they observed the same relationship at the single-patient level, although with lower decoding accuracies; they have also observed a similar relationship between dataset duration and prediction accuracy.
&nbsp;
As for whether the anatomical locations of the implanted electrodes have any implications on reconstructing music, the researchers found:
1.bilateral STG represented unique musical information compared to other regions;
2.right STG had unique information compared to left STG;
3.part of the musical information in left STG was redundantly encoded in right STG.
&nbsp;
As for the implications of the functional elements of responsive electrodes on reconstructuring music, the researchers found:
1.right onset electrodes had unique information compared to left onset electrodes and that part of the musical information in left onset electrodes was redundantly encoded in right onset electrodes. A similar pattern of higher right hemisphere involvement was observed with the late onset component;
2.right rhythmic electrodes had unique information, none of which was redundantly encoded in left rhythmic electrodes;
3.despite the substantial number of sustained electrodes, no impact of removing any set was found.
&nbsp;
<h3>Why is it important to reconstruct music from the brain?</h3>
&nbsp;
In fact, this type of research can help us better understand how music and language are processed. Also, it can give us some empirical insights into relevant diseases. For instance, similar research may reveal why people with Broca&#8217;s aphasia can usually sing the same words that they struggle to speak without difficulty.
&nbsp;
The study also lays the foundation for endowing synthesized speech with emotions. The researchers believe that the results will be very helpful for brainwave-based speech synthesis. Researcher Bellier hopes that this model will improve brain-computer interface technology so that speech aids will be able to reconstruct not only the speech itself, but also the intent expressed in conversational rhymes.
<a href="https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002176">Read the paper on Plos Biology</a>

[:en]Recently, scientists at UC Berkeley have succeeded in reconstructing the music the brain hears from brain waves. The results came from UC Berkeley's Helen Wills Neuroscience Laboratory, and Professor Gerwin Schalk, Director of the Chen Frontier Lab for Applied Neurotechnology, was one of the researchers.

&nbsp;

A total of 29 patients with drug-resistant epilepsy who had undergone intracranial electrode implantation to monitor seizures were included in this study, with a total of 2,668 electrodes implanted. The researchers asked the patients to listen to a certain piece of music and thereafter used artificial intelligence to decode and reconstruct the information recorded by the electrodes.

&nbsp;
<h3>Which brain regions have registered music?</h3>
&nbsp;

To identify electrodes encoding acoustical information about the song, the researchers used artificial intelligence to fit spectrotemporal receptive fields for all 2,379 artifact-free electrodes in the dataset, assessing how well the high-frequency activity (HFA) recorded at these sites could be linearly predicted from the song’s auditory spectrogram: the better the prediction, the more the electrodes at that location are associated with music recording.

&nbsp;
<h3>How much information is required to reconstruct musi from the brain?</h3>
&nbsp;

The researchers found that 80% of the best prediction accuracy was obtained with 43 electrodes; they observed the same relationship at the single-patient level, although with lower decoding accuracies; they have also observed a similar relationship between dataset duration and prediction accuracy.

&nbsp;

As for whether the anatomical locations of the implanted electrodes have any implications on reconstructing music, the researchers found:

1.bilateral STG represented unique musical information compared to other regions;

2.right STG had unique information compared to left STG;

3.part of the musical information in left STG was redundantly encoded in right STG.

&nbsp;

As for the implications of the functional elements of responsive electrodes on reconstructuring music, the researchers found:

1.right onset electrodes had unique information compared to left onset electrodes and that part of the musical information in left onset electrodes was redundantly encoded in right onset electrodes. A similar pattern of higher right hemisphere involvement was observed with the late onset component;

2.right rhythmic electrodes had unique information, none of which was redundantly encoded in left rhythmic electrodes;

3.despite the substantial number of sustained electrodes, no impact of removing any set was found.

&nbsp;
<h3>Why is it important to reconstruct music from the brain?</h3>
&nbsp;

In fact, this type of research can help us better understand how music and language are processed. Also, it can give us some empirical insights into relevant diseases. For instance, similar research may reveal why people with Broca's aphasia can usually sing the same words that they struggle to speak without difficulty.

&nbsp;

The study also lays the foundation for endowing synthesized speech with emotions. The researchers believe that the results will be very helpful for brainwave-based speech synthesis. Researcher Bellier hopes that this model will improve brain-computer interface technology so that speech aids will be able to reconstruct not only the speech itself, but also the intent expressed in conversational rhymes.

<a href="https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002176">Read the paper on Plos Biology</a>[:zh]音乐是生命体验中不可或缺的一部分。过去，人们想要表达音乐只能凭歌喉哼唱、靠乐器演奏；而如今，科学家已经能通过解码大脑来呈现脑中音乐——近期，加州大学伯克利分校（University of California, Berkeley，UC Berkeley）的科学家们成功地从脑电波中重建了大脑所听到的音乐。

&nbsp;

这项成果来自UC Berkeley的Helen Wills神经科学实验室，于今年8月15日发表于Plos Biology杂志，天桥脑科学研究院（Tianqiao and Chrissy Chen Institute，TCCI）应用神经技术前沿实验室主任Gerwin Schalk教授也是研究者之一。此前，该团队已经成功通过脑电波重建脑内语音。而这次的研究更进了一步，音乐所包含的信息显然远远大于语音。正如研究团队成员介绍，“音乐本质上是充满情感和韵律的——它有节奏、重音、抑扬顿挫，包含了比任何语言中有限的音素更广泛的含义。”

&nbsp;

有趣的是，与使用古典音乐的传统方式不同，研究人员重建的音乐片段是来自英国摇滚乐队Pink Floyd发表于1979年的歌曲“Another Brick in the Wall, Part 1”。为什么团队选择了Pink Floyd的音乐、特别是这个片段呢？“在论文中，我们提到的科学原因是：这首歌非常具有层次感，它引入了复杂的和弦、不同的乐器和不同的节奏，使得分析变得有趣。”认知神经科学家、该研究的主要作者Ludovic Bellier说道。“不过，不太科学的原因是我们真的很喜欢Pink Floyd。”

&nbsp;

这项研究共纳入了29名耐药性癫痫患者，他们均接受过颅内电极植入，以监测癫痫的发作。纳入本次研究的电极共2668个，每位患者36个到250个不等。研究人员为患者播放了“Another Brick in the Wall, Part 1”，让他们聆听这段音乐，此后利用人工智能对电极记录的信息进行解码、重建。

&nbsp;

先来听听这惊人一致的原曲和脑电波重建曲吧！可以听到，虽然重建的音频有些模糊不清、仿佛“在水下说话”，但乐曲的走向、一些重音处的歌词如“the wall”以及节奏都是清晰可辨的。

&nbsp;
<h3>哪些脑区记录了音乐？</h3>
&nbsp;

为了确定哪些部位的电极记录了歌曲声学信息的编码，研究者利用人工智能对2379个无伪迹电极记录的脑波信息进行了谱时感受野（STRF）拟合，评估不同位置电极记录的神经元高频活动到底能够多准确地被歌曲的听觉频谱图预测：预测程度越好，则该位置的电极便与记录音乐越相关。

[caption id="attachment_754" align="aligncenter" width="600"]<img class="wp-image-754" src="https://admin.next-question.com/wp-content/uploads/2023/11/50.png" alt="" width="600" height="735" /> ▷图1：响应电极的解剖位置[/caption]

&nbsp;

图1A表示所有电极的覆盖范围。图1B显示，347个电极具有显著的STRF拟合结果，位于左半球的有199个，右半球的有148个。这347个响应电极绝大多数（87%）集中在三个区域：68%位于双侧颞上沟（STG），14.4%位于双侧感觉运动皮层（SMC，位于中央前回和中央后回），4.6%位于双侧额下回（IFG）。图1C、D中，较深的颜色表示电极位于右半球，较浅的表示位于左半球；双因素ANOVA分析显示，两侧半球的对比有统计学意义，电极响应更加集中的区域均是右半球。

&nbsp;
<h3>从脑海中重建歌曲需要多少信息？</h3>
&nbsp;

科学家们从347个响应电极中随机抽取电极记录的数据，再利用人工智能解码其中信息、进行歌曲的重建。研究者发现，随机使用43个电极的数据即可达到最佳准确预测能力的80%；在单个患者上也类似，43个电极的信息已经可以进行解码，尽管解码的准确性较低；使用数据的持续时间与预测准确性之间也存在类似关系，例如，相比于使用完整的190.72秒的歌曲数据，使用69秒的数据即可以获得90%的重建准确性。

&nbsp;

那么，放置电极的解剖位置对重建是否有影响呢？在移除不同解剖位对的电极信息后再解码，发现：

（1）相比其他脑区，双侧STG具有独特的音乐信息；

（2）相比左侧STG，右侧STG具有独特的信息；

（3）左侧STG的部分音乐信息在右侧STG存在冗余编码。

[caption id="attachment_755" align="aligncenter" width="600"]<img class="wp-image-755" src="https://admin.next-question.com/wp-content/uploads/2023/11/51.png" alt="" width="600" height="322" /> ▷图2：不同音素的解码[/caption]

&nbsp;

不同解剖位置的电极在解码音乐时是否具有不一样的功能？确实如此。在对所有响应电极的独立成分分析后，结果如图2所示：

&nbsp;

- 仅位于双侧后STG的“起始成分”，记录主音吉他或合成器的起始部分、及人声中音节核心的起始部分（图2B、C、D的第一行）；

&nbsp;

- 位于双侧中、前STG以及双侧SMC的“持续成分”，记录歌曲的人声部分（图2B、C、D的第二行）；

&nbsp;

- 位于双侧后、前STG，以及双侧SMC的“迟发型起始成分”，也与主音吉他或合成器的起始部分、及人声中音节核心相关，只是潜伏期更长（图2B、C、D的第三行）；

&nbsp;

- 位于双侧中STG的“节奏成分”，记录歌曲中速度为99bpm、贯穿整个歌曲的节奏吉他中的16分音符（图2E）。

&nbsp;

了解了大脑是如何“接收”与“理解”音乐信号后，再来看响应电极的功能成分对音乐重建又有哪些影响？在移除不同相关功能的响应电极后再解码，发现：

&nbsp;

- 右侧起始电极相比左侧具有独特的信息；左侧起始电极的部分信息在右侧起始电极中存在冗余编码。对于迟发性起始电极也观察到类似现象。

&nbsp;

- 右侧节奏成分电极具有独特的信息，没有任何信息在左侧节奏电极中存在冗余编码。

&nbsp;

- 尽管持续电极数量很多，但移除它们的信息未发现任何影响。

&nbsp;

不过，由于电极的功能成分存在一定重叠性，所以对它们功能重要性的解读没有解剖位置那么准确。

&nbsp;
<h3>为什么关注大脑中重建的音乐？</h3>
&nbsp;

或许有人会对这项研究的目的产生疑问，为何我们要去听大脑对外界声音的映射？

&nbsp;

其实，这类研究能够帮助我们更好地理解音乐和语言的处理方式。另外，在对疾病的认识上也能给我们一些实证性的启发，例如揭示为什么布罗卡失语症患者讲话费力，但通常却可以毫无困难地用唱歌的方式唱出相同的词。

&nbsp;

这项研究也为将情感赋予合成语音奠定了基础。尽管研究重点放在音乐上，但研究人员认为，这项结果对于基于脑电波的语音合成将有很大帮助。无论哪种语言，人类的话语都包含着节奏、重音、抑扬顿挫等音乐性要素，这些要素构成了话语之中隐含的情感成分。“这些元素，我们称之为韵律，携带着无法仅仅用语言表达的意义。”研究者Bellier希望这个模型能够改进脑机接口技术，使语言辅助不仅能重构语音本身，还能重构话语中用韵律表达的意图。

&nbsp;

过去，科学家们已成功重建脑内语音，能使中风或肌萎缩侧索硬化症等神经系统疾病患者通过植入式语音解码器来表达自己；但此类重建通常是机械、刻板的。研究者们希望这项成果最终能帮助失语患者恢复自然言语的音乐性。

&nbsp;

“如果脑机接口能够用音乐中固有的韵律和情感来重新创造某人的言语，那么它所能重建的不仅仅是单词，不是机械地说，‘我，爱，你，’而是可以像真人一样大喊，‘我爱你！’”

&nbsp;

另外，这项研究重建的是研究对象听到的音乐，而研究者Robert Knight认为，未来的研究应向重建脑中想象的语音、音乐发展。“虽然他们没有记录受试者想象音乐时的大脑反应，但这可能是脑机接口未来的用途之一：将想象的音乐转化为真实的音乐。”

&nbsp;

不过，这项技术离实际运用还有很长的路要走，比如，目前获取的数据还基于有创的颅内植入电极；或许未来可以利用无创电极来实现数据收集，这将建立在脑机接口技术的发展之上。

&nbsp;

总之，一旦这项技术成熟应用，是否我们就能依靠一个轻型头盔来创作心中的音乐呢？那时候会有怎样瑰丽的想象化为现实？让我们无限遐想。

&nbsp;

<a href="https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002176">阅读原文</a>[:]

This research recorded the music heard by your brain!

[:en]This research recorded the music heard by your brain![:zh]你听到的音乐在脑中是怎样的？这项研究把它录下来了！[:]

Recently, a team of researchers from Europe published a study entitled Is Short Sleep Bad for the Brain? Brain Structure and Cognitive Function in Short-Sleepers in the Journal of Neuroscience. Their research results indicate: daytime sleepiness and sleep problems are more strongly related to regional brain volumes than sleep duration. Those with short sleep duration had slightly lower scores on tests of general cognitive function (GCA). The association between habitual short sleep and lower scores on tests of general cognitive abilities must be further scrutinized in natural settings.
&nbsp;
<h3>Is it really unhealthy to sleep fewer hours than recommended?</h3>
&nbsp;
It is often believed that reduced sleep can damage brain health and affect the cognitive function of the brain. However, you may find that you are surrounded by people who sleep less than the recommended amount of time and are not sleepy during the day.
&nbsp;
This may be attributed to two factors:
First, partly due to genetic differences (Dashti et al. 2019), sleep need varies.
Second, like other physiological systems, sleep need may not be fixed, but is expected to have the ability to adapt to changes in external circumstances (Horne 2011).
&nbsp;
A subset of healthy adults sleeping seven-to-eight hours without daytime
sleepiness can still fall asleep within 6 minutes in the multiple sleep latency test (MSLT). This indicates a level of sleepiness similar to patients with primary sleep disorders, and it was suggested that this could represent accumulated sleep debt from chronically mild insufficient sleep. On the other hand, duration per se may be a less important indicator of insufficient sleep than daytime sleepiness and problems such as frequently having trouble falling or staying asleep.
&nbsp;
<h3>Is sleep duration related to brain volumes and cognitive functions?</h3>
&nbsp;
Researchers explored the relationship between sleep duration and brain volume and cognitive function and found that brain volume was greater with short sleep duration and without sleep problems and daytime sleepiness; however, those with short sleep duration had slightly lower cognitive performance.
&nbsp;
First of all, the researchers classified participants in four groups: 
Group 1: short sleep duration, no sleep problems, no daytime sleepiness;
Group 2: short sleep duration, sleep problems, daytime sleepiness;
Group 3: normal sleep duration, no sleep problems, no daytime sleepiness;
Group 4: normal sleep duration, sleep problems, daytime sleepiness.
&nbsp;
By comparing the brain volumes of different groups, the researchers found that Group 2 had significantly smaller brain volumes than Group 1. The differences were most evident for the brain stem and pallidum. Group 1 had larger brain volumes than Group 3, driven by differences in brain stem, cerebellum white matter and caudate volumes. Group 4 tended to show smaller volumes for cerebellum WM, corpus callosum and pallidum compared with Group 3. However, there were no significant overall differences across regions. More analysis revealved that there were no substantial differences in brain volumes between Group 2 and Group 4 neither.
&nbsp;
Next, they assessed the cognitive function differences between different groups via GCA. It turned out that There were no significant differences between the groups of short sleepers (Group 2 vs Group 1). Group 4 showed significantly lower GCA than Group 3. Both Group 1 and Group 2 showed significantly lower scores than Group 3.
&nbsp;
In conclusion, individual sleep needs vary as affected by genetic and environmental factors. Perhaps the best choice is to take an uninterrupted sleep when you feel sleepy.
<a href="https://www.jneurosci.org/content/43/28/5241">Read the paper on Journal of Neuroscience</a>

[:en]Recently, a team of researchers from Europe published a study entitled Is Short Sleep Bad for the Brain? Brain Structure and Cognitive Function in Short-Sleepers in the Journal of Neuroscience. Their research results indicate: daytime sleepiness and sleep problems are more strongly related to regional brain volumes than sleep duration. Those with short sleep duration had slightly lower scores on tests of general cognitive function (GCA). The association between habitual short sleep and lower scores on tests of general cognitive abilities must be further scrutinized in natural settings.

&nbsp;
<h3>Is it really unhealthy to sleep fewer hours than recommended?</h3>
&nbsp;

It is often believed that reduced sleep can damage brain health and affect the cognitive function of the brain. However, you may find that you are surrounded by people who sleep less than the recommended amount of time and are not sleepy during the day.

&nbsp;

This may be attributed to two factors:

First, partly due to genetic differences (Dashti et al. 2019), sleep need varies.

Second, like other physiological systems, sleep need may not be fixed, but is expected to have the ability to adapt to changes in external circumstances (Horne 2011).

&nbsp;

A subset of healthy adults sleeping seven-to-eight hours without daytime

sleepiness can still fall asleep within 6 minutes in the multiple sleep latency test (MSLT). This indicates a level of sleepiness similar to patients with primary sleep disorders, and it was suggested that this could represent accumulated sleep debt from chronically mild insufficient sleep. On the other hand, duration per se may be a less important indicator of insufficient sleep than daytime sleepiness and problems such as frequently having trouble falling or staying asleep.

&nbsp;
<h3>Is sleep duration related to brain volumes and cognitive functions?</h3>
&nbsp;

Researchers explored the relationship between sleep duration and brain volume and cognitive function and found that brain volume was greater with short sleep duration and without sleep problems and daytime sleepiness; however, those with short sleep duration had slightly lower cognitive performance.

&nbsp;

First of all, the researchers classified participants in four groups:
Group 1: short sleep duration, no sleep problems, no daytime sleepiness;

Group 2: short sleep duration, sleep problems, daytime sleepiness;

Group 3: normal sleep duration, no sleep problems, no daytime sleepiness;

Group 4: normal sleep duration, sleep problems, daytime sleepiness.

&nbsp;

By comparing the brain volumes of different groups, the researchers found that Group 2 had significantly smaller brain volumes than Group 1. The differences were most evident for the brain stem and pallidum. Group 1 had larger brain volumes than Group 3, driven by differences in brain stem, cerebellum white matter and caudate volumes. Group 4 tended to show smaller volumes for cerebellum WM, corpus callosum and pallidum compared with Group 3. However, there were no significant overall differences across regions. More analysis revealved that there were no substantial differences in brain volumes between Group 2 and Group 4 neither.

&nbsp;

Next, they assessed the cognitive function differences between different groups via GCA. It turned out that There were no significant differences between the groups of short sleepers (Group 2 vs Group 1). Group 4 showed significantly lower GCA than Group 3. Both Group 1 and Group 2 showed significantly lower scores than Group 3.

&nbsp;

In conclusion, individual sleep needs vary as affected by genetic and environmental factors. Perhaps the best choice is to take an uninterrupted sleep when you feel sleepy.

<a href="https://www.jneurosci.org/content/43/28/5241">Read the paper on Journal of Neuroscience</a>[:zh]每次熬夜的时候都在想，如果我们不需要睡眠却又精力充沛，那该是一件多么美好的事情。

&nbsp;

正常人熬夜（睡眠时间少于6小时）后，白天就会出现精神萎靡的情况，想要补觉。但也存在这样一类人，他们的睡眠时间少于6小时，但白天却总是精神饱满，一点也看不出疲态。难道他们是“天选之子”？为什么他们需要的睡眠时间比正常人短？他们的大脑有何特殊之处？

&nbsp;

近日，来自欧洲的多个研究团队一同在Journal of Neuroscience发表题为“Is Short Sleep Bad for the Brain? Brain Structure and Cognitive Function in Short Sleepers”的研究，结果表明：与睡眠时长相比，特定脑区的体积与白天嗜睡和睡眠障碍的相关性更强。睡眠时间短的人，在一般认知功能测试（GCA）上的得分仅仅略低。至于睡眠时间短是否会导致认知能力下降，不能仅通过GCA得分确定，还有必要在现实环境中进一步验证。

&nbsp;

一般认知功能测试（General Cognitive Function Test，缩写为GCA）是一种用于评估个体普遍认知能力的测试方法。这种测试旨在测量人们在各种认知任务中的表现，包括注意力、记忆、思维能力、语言理解、问题解决和推理等方面。通过评估这些认知领域的综合表现，一般认知功能测试可以提供关于一个人的智力水平和认知健康状况的信息。这种测试对于研究认知障碍、大脑健康以及认知能力在不同人群中的差异等方面具有重要意义。

&nbsp;

[caption id="attachment_744" align="aligncenter" width="600"]<img class="wp-image-744" src="https://admin.next-question.com/wp-content/uploads/2023/11/41.png" alt="" width="600" height="124" /> ▷图源：Journal of Neuroscience官网[/caption]

&nbsp;
<h3>睡得少于推荐睡眠时间就不健康吗？</h3>
&nbsp;

据统计，大约有50%的人睡眠时间少于推荐睡眠时间（7-9小时），其中更有6.5%睡眠时间少于6小时。通常人们认为，睡眠减少会损害大脑健康，影响大脑认知功能。然而，在生活中你可能会发现，身边确有许多睡得比推荐睡眠时间少，白天却不嗜睡的人。

&nbsp;

这种现象可能来源于两个因素：

第一，基因水平上的差异导致了个体水平上的睡眠需求不同。

第二，类似于其他的生理稳态系统，睡眠需求并不是固定不变的，而是能随着环境变化而动态变化的。

&nbsp;

这两点使得睡眠时间短，却不出现睡眠障碍以及白天嗜睡成为可能。当然，睡眠时间如何影响日间精神状态和认知功能，仍然是一个亟待解决的复杂问题。

&nbsp;

在多重睡眠潜伏期测试中，一些睡眠时间正常（7-8小时）并且没有出现白天嗜睡以及睡眠障碍的成年人却在6分钟内快速入睡，这一现象表明这些人与原发性睡眠障碍的患者类似，背上了长期轻度睡眠不足导致的积累性睡眠债务。在长时间的轻微睡眠剥夺后，精神萎靡的恢复比精神警觉任务表现水平恢复快。此外，测试还发现，如果睡眠时间短的原因是更低的睡眠需要，那么其实不大可能影响认知功能和大脑健康。

&nbsp;

多重睡眠潜伏期测试（Multiple Sleep Latency Test，简称MSLT）是一种用于评估睡眠质量和昼夜节律的医学检测方法。在测试中，被试在一系列特定时间段内被要求躺下休息，以观察他们在不同时段内入睡的速度，从而评估他们的睡眠倾向和白天的困倦程度。这种测试通常用于诊断睡眠障碍，如嗜睡症和睡眠呼吸暂停症等。

&nbsp;

精神警觉任务（Sustained Attention Task）是一种用于评估个体持续注意力水平和集中精力能力的认知任务。在这种任务中，被试需要在一段较长的时间内保持对特定刺激或任务的关注，并根据特定规则执行操作。这种任务要求被试在一段时间内保持警觉，以检测可能的变化或目标。常见的精神警觉任务包括连续性性能测试（Continuous Performance Test）和选择性持续性性能测试（Selective Attention Sustained Attention Task）等。通过精神警觉任务，研究人员可以了解被试在持续时间较长的任务中的注意力表现，以及他们是否能够在可能的干扰下保持警觉。这种任务在评估认知功能、注意力缺陷和注意力障碍等方面具有重要作用，对于研究个体的注意力特点以及注意力问题的诊断和干预都很有价值。

&nbsp;

另一方面，测试也发现，单纯依靠睡眠时间可能很难判断一个人是否睡眠不足，而应该结合是否出现白天嗜睡、入睡困难或难以保持睡眠状态等多种睡眠障碍现象。白天日嗜睡与更薄的皮层有关，但与睡眠时长是否有关，还有待研究。

&nbsp;
<h3>睡眠时长与大脑体积及认知功能有关吗？</h3>
&nbsp;

研究人员利用HCP、UKB、BASE-II等多个项目的数据，探究了睡眠时长与大脑体积及认知功能的关系，发现睡眠时间短且无睡眠问题及白日嗜睡的情况下，大脑体积更大；但睡眠时间短的人，其认知能力略低。

&nbsp;

首先，研究者将数据分为四组：

1.睡眠时间短，无睡眠问题，无白日嗜睡；

2.睡眠时间短，有睡眠问题，白日嗜睡；

3.睡眠时间正常，无睡眠问题，无白日嗜睡；

4.睡眠时间正常，有睡眠问题，白日嗜睡。

&nbsp;

随后，他们比较了各组之间的脑体积，发现第二组的脑体积明显小于第一组。而这种差异在脑干和苍白球最为明显。第一组的脑体积也要大于第三组，这种差异在脑干、尾状核、小脑白质和皮质体积中最为明显。第四组与第三组相比，其小脑白质、胼胝体和苍白球的体积较小，但在大脑整体上没有显著性的差异。他们还比较了第二组和第四组的脑体积，发现并没有显著性的差别。

<img class="aligncenter wp-image-745" src="https://admin.next-question.com/wp-content/uploads/2023/11/42.png" alt="" width="600" height="496" />

[caption id="attachment_746" align="aligncenter" width="600"]<img class="wp-image-746" src="https://admin.next-question.com/wp-content/uploads/2023/11/43.png" alt="" width="600" height="517" /> ▷图注：左，一二组之间各脑区体积的比较；右，三四组之间各脑区体积的比较[/caption]

&nbsp;

接着，他们通过GCA来比较各组之间的认知功能差异。在睡眠时间短的两组之间（第一组和第二组）认知功能没有显著性的差别。而第四组的认知能力显著低于第三组。而第一组和第二组的认知能力测试都低于第三组。

&nbsp;

研究者对各组间的BMI、抑郁程度、受教育水平的进一步分析发现，各变量在各组之间都有显著性的差异。与第一组相比，第二组有着更高的BMI，更高抑郁程度，以及更低的受教育水平。与第一组相比，第三组有着更低的BMI。与第一组相比，第四组有着更多的抑郁水平。

[caption id="attachment_747" align="aligncenter" width="600"]<img class="wp-image-747" src="https://admin.next-question.com/wp-content/uploads/2023/11/44.png" alt="" width="600" height="318" /> ▷图注：各组间BMI值、抑郁程度、受教育情况的比较[/caption]

&nbsp;

为了控制以上变量对实验结果的影响，他们重做了主要的分析（第一组和第二组；第三组和第四组），发现并不影响主要的结论，尽管各组间的BMI、抑郁程度、受教育水平有所不同，但是并不能足以解释各组间大脑体积的差异。

<img class="aligncenter wp-image-748" src="https://admin.next-question.com/wp-content/uploads/2023/11/45.png" alt="" width="600" height="330" />

[caption id="attachment_749" align="aligncenter" width="600"]<img class="wp-image-749" src="https://admin.next-question.com/wp-content/uploads/2023/11/46.png" alt="" width="600" height="268" /> ▷图注：控制BMI、抑郁程度、受教育水平后，第一组和第二组间、第三组和第四组间大脑各脑区体积的差异[/caption]

&nbsp;
<h3>结语</h3>
&nbsp;

脑形态测量和一般认知能力测试被用作衡量大脑健康和认知功能的指标，那么用其他指标衡量大脑健康和认知功能是否能够得到不同的结果呢？如果参与人员进行全面的筛查以排除睡眠呼吸暂停等睡眠障碍，这是否会影响到所观察到的睡眠时间与大脑的关系？由于遗传因素和环境因素，个体的睡眠需求有很大差异，也许最好的选择是瞌睡来了就睡，一睡睡到自然醒。

&nbsp;

<a href="https://www.jneurosci.org/content/43/28/5241">阅读原文</a>[:]