In life, we have been memorizing, sequencing, and flexibly calling all kinds of information, such as thinking about the order of cards when playing cards, planning routes when traveling, and arranging the order of work content.
Ordering is so frequent and natural that people often don't realize how complex these tasks are. If a computer is used to achieve this kind of sorting task, it is necessary to encode, store, calculate, compare, and call a series of tedious steps to encode, store, calculate, compare, and call the information through the cooperation of hardware and algorithms. So how does the brain accomplish this function?
不同於計算機的硬體結構，大腦是由上億個神經元細胞組成的龐大網路。人們雖然瞭解它的種種認知功能，卻對其底層邏輯所知甚少，如同一個“黑盒子”。近日，一項研究通過對獼猴進行心理排序任務時神經元活動的記錄和分析，揭示了大腦的工作記憶模式。該研究題為《獼猴額葉皮層對空間序列資訊的工作記憶程式設計》，於2024年9月27日發表在《科學》（Science）雜誌上，作者是來自中國科學院腦科學與智慧技術卓越創新中心（以下簡稱“腦智卓越中心”）的王立平團隊。
In this study, two macaques were trained to perform a delayed sequence sequencing task. They need to memorize the dots that appear on the screen at different sequences, then sort the sequences forward or backward when prompted, and finally report the results by touching the screen. The researchers implanted micro-drive electrode arrays into the prefrontal cortex of macaques, recorded the electrical activity of thousands of neurons, and mathematically described the results to analyze the patterns that emerged.
The study found that in the sequencing task, the information of each order is recorded in a corresponding "subspace" that reflects the different global states of neurons. When faced with cognitive tasks that require reversing order, sequential information in subspaces is exchanged by storing and transferring information by forming temporary subspaces. In addition, when macaques are asked to invoke order information in different rules in order or reverse order, there is a subspace that stores the rules to control the flow of information.
The study deepens the understanding of the underlying logic of the brain's working memory. Using these spatial patterns, researchers were even able to infer what macaques saw and what they were ranked in reverse, as if they were "mind reading."
What is the difference between macaques and humans' sorting abilities? How do we understand the brain laws that these "subspaces" represent? How to study more complex cognitive phenomena such as "consciousness"? In order to answer these questions, The Paper recently interviewed Tian Zhenghe, the first author of the study and a doctoral student at the Center for Excellence in Brain, and Liping Wang, the corresponding author and researcher at the Center for Excellence in Intelligence.
From Function to Bottom: Opening the "Black Box" of Working Memory
In neurocognitive science, ordering things in the mind is thought to be closely related to working memory. Unlike long-term memory, working memory mainly involves the maintenance and longitudinal management of short-term memory in response to the cognitive situation that needs to be faced immediately.
"A lot of theories arise from some intuitive experience, and then they are constantly being revised." Wang Liping said, "The difference between long-term and short-term memory is quite intuitive - some things can be remembered for a long time, while others are quickly forgotten. Later, it was discovered that memories are 'longitudinal' in the brain, such as 'compressed': many memories are missing details and only impressions. In short-term memory, this longitudinal is embodied in solving the problem of the moment by combining memories of limited capacity. If there is only 'memory' and no 'work', then nothing can be done. ”
Wang Liping said that the longitudinal part of working memory involves cognitive processes such as reasoning, which is relatively complex, and there are still few related studies. This study aims to fill this gap.
Whether it is the storage or production of memories, it is necessary to find the corresponding process at the physical level of the brain and verify it to form a complete explanation. Structurally and morphologically, the brain is roughly divided into four parts: frontal, parietal, temporal and occipital lobes. Previous studies have found that neurons in the prefrontal cortex exhibit persistent activity in working memory tasks and are able to send signals from top to bottom that affect other areas of the brain.
而在細胞層面，大腦是由上億個（人類為860億個）不同種類的神經元細胞通過突觸連接而成的複雜網路，通過電流等信號遞質進行交流。從這個角度上來看，大腦並不是像電腦那樣以功能明確的模組化方式運行的，而是一個大規模的並行網路：同樣的神經元可能廣泛參與不同的認知活動。
隨著技術的發展，科學家們已經能夠同時監測更多單個神經元的活動，從底層還原大腦的運行狀態。在這項研究中，研究者們使用了157通道的微電極陣列同時測量了獼猴前額葉皮質中4191個神經元在工作記憶任務過程中的活動狀況。
With the technical means to monitor neuronal activity, researchers also need to design appropriate experiments to ensure that the measurements and analyses can lead to strong interpretations. Wang Liping believes that although working memory involves complex cognitive functions, it can also be broken down into clear and simple processes to illustrate.
"True intelligence is abstract. But the representations of high-level cognitive functions that are considered complex in the brain, such as reasoning, doing math, using symbols, etc., can be studied. As long as we find the right way to break down and combine the questions, we can try to answer more abstract questions, such as those about consciousness. He said.
In this study, we broke down the mental sequencing into three stages through an ingenious experiment. In the first stage, the macaque needs to remember 700 to 0 dots on the screen that flash sequentially. After a short delay (0 to 0 milliseconds), in the second stage, the macaque will see an image, and if it sees a cucumber, it means that it needs to be debriefed in order. If you see Apple, it needs to report the order backwards.
In this way, the macaque stores sequential memories in the brain and carries out forward or reverse longitudinal longitudinal on these memories. Also after a short delay, in the final stage, it is prompted by a picture (a blue dot) to start the presentation, followed by a touch of the screen to tap on the result.
The macaques used in the study were trained over a long period of time, allowing them to focus on the task and to be able to understand the meaning of the picture instructions. Tian Zhenghe told The Paper that the memory ordering and even understanding symbols are mastered by many animals, and are not "advanced" activities unique to humans. In this sense, macaque experiments can be an important reference for exploring the laws of the human brain.
Characterizing the brain: Describe how neural networks operate
After measuring the macaque's neuronal activity during the sequencing task, the next step is to analyze the data to see what the brain is "doing" in the process.
When faced with memory and sequencing tasks, neurons in the macaque's brain are active at the same time, and the firing intensity of each neuron varies from high to low, forming different states of the brain. Wang Liping said that it is like a symphony orchestra, although it is playing a piece, the melody, rhythm and strength of the different members are different. By grouping individual members into one category, such as "String Ensemble" and "Wind Ensemble", you can "peek into the leopard" and explore the pattern of the entire repertoire.
The investigators first wanted to establish a link between the task stimulus and the neuronal state. It can be understood as establishing an equation for the relationship between the state of a single neuron (the dependent variable y) and the order information (independent variable x), where the state of the neuron changes with the order information faced by the macaque. We measured more than 4000 neurons, and there were more than 0 such equations. Tian Zhenghe said that he studied condensed matter physics as an undergraduate, and he has strong mathematical thinking.
"The solution of each set of equations represents the size of the response of a certain neuron to the information of each order and position, so the solution of 4000 sets together is the group response of all neurons to these positions." "Let's go to the most salient directions, or patterns, of these groups' responses," he said. ”
From a vector point of view, the overall state of these neurons at a given moment in time is a point in a multidimensional space of 4000. In the face of different tasks and at different times, the points representing these states form a geometric structure in this high-dimensional space. Using mathematical methods such as principal component analysis (PCA), the research team was able to identify "subspaces" with overarching characteristics that reflect information about a specific task.
Tian Zhenghe explained that mathematically speaking, the subspace represents which of those more than 4000 equations is more significant in the face of a specific task. At the brain level, the subspace represents a certain combination of all these neurons, each of which plays a role.
Since subspaces reflect specific patterns, it can be understood that these spaces "record" the information stored by the brain. The research team found that information about the positions of multiple dots on the screen was stored in different subspaces, and that this information remained stable over time when the macaques did not need to sort the dots.
When the macaque sees the apple picture and needs to rearrange the order, each original subspace will recruit a new temporary subspace, first pass the original internal memory information to it, and then pass the memory signal in the temporary subspace to the space to be exchanged after its own information is emptied.
"It's like exchanging water in a cup. Usually we exchange water in two cups and just get a new cup. And for the brain, it will take out two new cups to exchange. Tian Zhenghe said.
The study also found that there is a subspace that does not record specific order information, but is related to the rules of forward or reverse order. Under different rules, the state of this subspace will show different development trajectories over time. The research team speculated that this subspace controls the flow of information between the sequential and temporary subspaces, and can initiate and gated exchange processes.
Tian Zhenghe explained that if the metaphor of exchanging water in a cup is followed, then this subspace is like a waiter carrying water, and the guest to his left needs to exchange the water in the cup, while the guest to his right does not. When the guest on the left greets him, he walks over to that side and pulls out two new cups. When the guest on the right greets him, he will bring two glasses of water to him.
"Of course, the water, the cups, the waiters, and the exchange of motions here are all different states of the brain, like different sides of a face." Tian Zhenghe said.
With the improvement of technical means, scientists can obtain more and more biological data, and it is becoming more and more low-level. In order to characterize and analyze complex biological systems such as the brain, in addition to biological knowledge, it is increasingly important to collaborate extensively with disciplines such as mathematics, computer science, physics, chemistry, etc. Wang Liping mentioned that many of the doctoral students recruited by his research team are from other majors, and they have brought many new research ideas through the collision of ideas from different perspectives.
"In many fields, tutors don't necessarily know more than students, and they need to learn from each other." He said.
As the first author of the study, Tian Zhenghe believes that the atmosphere of collaboration and open discussion in the research group is very important for conducting breakthrough research, "It is very important to be able to express and experiment with your ideas freely." ”
(Reporter Ji Jingjie)