Sohu Technology's "Big Bang of Ideas - Dialogue with Scientists" column No. 15, in conversation with Wu Si, Dean of the School of Psychological and Cognitive Sciences of Peking University.
Guest profiles
Wu Si is a professor and dean of the School of Psychological and Cognitive Sciences of Peking University, a researcher at the McGovern Institute of Brain Science, a researcher at the Center for Quantitative Biology, a researcher at the Peking University-Tsinghua Joint Center for Life Sciences, and a scholar in Beijing. His main research interests are computational neuroscience and brain-like computing, and he currently serves as the editor-in-chief of Frontiers in Computational Neuroscience and the director of the "Computational Neuroscience and Neural Engineering Professional Committee" of the Chinese Neuroscience Society.
Focus:
01
The volume of the brain is necessary, but not sufficient, for complex behavior. There are specialized brain areas responsible for emotions in the brain, and there is also the hippocampus responsible for memory, if you want artificial intelligence to really produce emotions and other advanced functions, the large size of the model is on the one hand, and on the other hand, it is necessary to design a good structure.
02
AI will most likely replace basic psychological counselors, and the reason why there are no relevant mature products is that the people who develop large models understand technology, but lack professional psychological knowledge background, and need talents in these two fields to cooperate with each other to make more humane and empathetic applications.
03
ChatGPT is just a language model, but not an agent, it is like a military master to give you ideas, but the execution still needs people to do, and the current AI has not yet achieved the so-called 'embodied intelligence'.
04
So far, we know very little about many basic locomotion mechanisms in humans, and people think that learning to ride and swim is very simple, but it may be more difficult than large models in terms of technical difficulty, which is one of the reasons why the current progress in the field of robotics is slow.
05
Some people may say that the current artificial intelligence is doing very well, Boston mechanical dogs can walk, turn their heads, we think it's too great, in fact, compared to animals in nature, a duck is born to run.
Produced | Sohu Technology
Author | Zheng Songyi
"Among the things and phenomena known in nature, humans and the human brain are the most complex systems, and human intelligence is the most complex phenomenon. Therefore, brain science has always been regarded as the 'last frontier of natural science', and the modern method of artificial intelligence is the source of power to open up territory. ”
Some people speculate that the intelligence of the brain is proportional to the size of the brain, and as the AI model gets bigger and bigger, it will also have emotions, thinking and other characteristics. Wu Si, dean of the School of Psychological and Cognitive Sciences at Peking University, said in a conversation with Sohu Technology, "The volume of the brain is a necessary condition for producing complex behavior, but it is not a sufficient condition. There are specialized brain areas responsible for emotions in the brain, and there is also the hippocampus responsible for memory, if you want artificial intelligence to really produce emotions and other advanced functions, the large size of the model is on the one hand, and on the other hand, it is necessary to design a good structure. ”
As an optimist, Wu Si believes that artificial intelligence will most likely produce emotions and thinking abilities like transformers in the future. "People sometimes overestimate the current level of technology, but also underestimate the level of technology in the future," he said. ”
Wu Si introduced that artificial intelligence and brain science are highly intersecting research fields, and deep learning networks are to imitate the basic structure of the brain's nervous system. He said that the close intersection between artificial intelligence and psychology is also recognized in the industry, and artificial intelligence algorithms provide basic tools for psychological research.
"The current demand for mental health counseling is particularly great in society, and the material level of our life is very good, but people are very confused mentally, including depression problems in adolescent students. However, if you train a psychological counselor according to the traditional way, the cost of time, manpower and financial resources is very large, and AI is a good tool to meet the needs of society.
He believes that AI will most likely replace basic psychological counselors, and the reason why there are no relevant mature products is that the people who develop large models understand technology, but lack professional psychological knowledge background, and need talents in these two fields to cooperate with each other to make more humane and empathetic applications.
Talking about the development status and future development direction of artificial intelligence, he said, "The technical route of imitating brain structure and realizing intelligence through neural network connections is to be adhered to." In addition to the vision system, hearing, smell, perception system, and motion control system will all wait for future AI realization. ”
"There is a saying that there are about 400 cognitive functions in humans, and current AI has achieved less than 10, such as the function of humans subconsciously processing and processing complex information. ChatGPT is just a language model, but not an agent, it is like a military master to give you ideas, but the execution still needs people to do, and the current AI has not yet achieved the so-called 'embodied intelligence'. ”
After the wind of AI large models blew, Tesla, Byte and many other large manufacturers shifted their business focus to AI robot research and development, but they were also frequently questioned by the outside world, "the long research and development cycle does not seem to have achieved technical breakthroughs, and it still maintains the traditional robot function model."
Wu Si explained, "So far, we have little understanding of many basic human movement mechanisms, people think that learning to ride and swim is very simple, but from the technical difficulty may be more difficult than large models, which is one of the reasons why the current progress in the field of robotics is slow." Some people may say that the current artificial intelligence is doing very well, Boston mechanical dogs can walk, turn their heads, we think it's too great, in fact, compared to animals in nature, a duck is born to run. So we have to figure these things out, equip artificial intelligence with the ability to execute, then we can look forward to the future of embodied intelligence, which is what everyone is doing. ”
The following is a transcript of the conversation (edited and edited)
Sohu Technology: Recently, people seem to be talking about "artificial intelligence", in your opinion, what does the essence of "intelligence" mean?
Wu Si: The essence of intelligence is a relatively complex problem, if from the performance level, it refers to the advanced cognitive abilities of human beings, including reasoning, learning, memory, the ability to solve complex problems, creativity, etc. From a computer science perspective, it means that we want to develop an artificial intelligence system to mimic human intellectual behavior.
The question of the nature of intelligence has puzzled us not only with us, but also with many scientists and philosophers for centuries. My current tendency to tend that the nature of intelligence comes from a supercomplex system that, in a sense, automatically emerges with incredible creative abilities, as well as advanced behavioral cognitive abilities.
Sohu Technology: Some people speculate that the intelligence of the brain is proportional to the volume, and the brain volume continues to increase in the process of human evolution over the past 14 million years, and gradually has the ability to emotion, consciousness, thinking, etc., so will the AI model also have emotions, consciousness and other abilities as it continues to increase?
Wu Si: It is possible that the size of the brain is a necessary but not sufficient condition for producing complex behavior. There are specialized brain areas responsible for emotions in the brain, and there is also the hippocampus responsible for memory, if you want artificial intelligence to really produce emotions and other advanced functions, the large size of the model is on the one hand, and on the other hand, it is necessary to design a good structure.
Sohu Technology: Knowing that you were a professor at the School of Information Science and Technology of Peking University, and later joined the School of Psychological and Cognitive Sciences, what is your original intention for the change of professional field?
Wu Si: Speaking of this original intention, I may have to go back a little further in history, I first had a background in physics, and during my Ph.D. I realized that many of the basic laws of physics were almost done by Einstein and others, but the basic laws of artificial intelligence and the human brain are not clear yet, so I changed my research direction to computational neuroscience during my postdoctoral period.
Computational neuroscience means that we use mathematical models to reveal the basic laws of the human brain, just as in physics, we use mathematical models to describe the laws of nature.
In the more than 20 years of scientific research after graduation, I have actually been doing computational neuroscience, the research direction is to reveal the basic working principle of the brain, and develop some artificial intelligence models on this basis. Therefore, my work in different units is completely in line with the needs of the times, for example, when I worked in the information department of Peking University, it was because the college needed some brain-like experts to cooperate with artificial intelligence experts. The current School of Psychology of Peking University wants to do cross-research between psychology and cognitive science and artificial intelligence, and my background is also very suitable for such an environment, so it is very natural to go to the School of Psychological and Cognitive Science of Peking University.
Sohu Technology: In your opinion, is there an interaction between brain science research and artificial intelligence research?
Wu Si: In fact, these two have been intersecting, such as the underlying technology of artificial intelligence, deep learning networks, which imitate the basic structure of our brain's nervous system.
When I started doing artificial intelligence research more than 20 years ago, it was a cold winter of artificial intelligence, without big data and super computing power, it was a bit desperate, so I began to study brain science and try to get inspiration from it. Now this field is called brain-like intelligence.
Now we will use AI models to explain the behavior of the brain when analyzing some nervous system working principles, and also summarize some models according to the working principle of the brain and develop brain-like intelligence algorithms, so the two fields have actually never been the same.
Sohu Technology: Now that you are the dean of the School of Psychological and Cognitive Sciences at Peking University, what do you think is the connection between psychological and cognitive science and artificial intelligence?
Wu Si: The close intersection of psychology and artificial intelligence has become a consensus in the industry, and the algorithm model of artificial intelligence provides basic tools for psychological research.
At present, the demand for mental health counseling in society is particularly large, and the material level of our life is very good, but people are very confused mentally, including depression problems in adolescent students. Our college will do some psychological counseling services, but currently faces a big problem, such as training a psychological counselor according to the traditional way, the cost of time, manpower and financial resources are very large, AI is a good tool to meet the needs of society.
In addition, when we study the neural mechanisms behind the activity of the human mind, the role of AI is also very large. First of all, why the progress of human brain research is relatively slow, the human brain will not be disassembled like a machine to see the working principle, so the research of the human brain uses some indirect methods to speculate brain activity through human behavior.
Artificial intelligence can provide a reference for biological intelligence research, and we train an artificial intelligence system to reveal how the biological brain works when it presents a certain cognitive function. For example, through comparison, it is possible to find that after a long day, some function that humans think is very advanced is actually so simple. For example, when the deep learning model does object recognition, we can see how the model does it, which can be compared with human vision, which provides a particularly good reference for us to understand the principle of biological vision.
Sohu Technology: As far as you know, are there AI psychological counselors in the current society? Do you think AI will replace the job of counselor?
Wu Si: AI will most likely replace basic psychological counselors, but there are no mature products yet, because the people who develop large models understand technology, but lack professional psychological knowledge background, and need talents in these two fields to cooperate with each other to make more humane and empathetic applications.
Sohu Technology: Some people say that the training of neural networks is to imitate brain networks, but we know that the brain nerves in brain networks are intricate, how can we summarize and extract the way the human brain learns and thinks, so as to find a way to optimize machine learning neural networks?
Wu Si: There are a large number of experimental neuroscientists in the field of brain science research, they use monkeys, mice, zebrafish, fruit flies and other animals to conduct research, because the brains of these animals can be turned on, some functions of their brains can be observed, and scientists will record a large number of experimental data. What we do as computational neuroscientists is to read the articles published by these experimental scientists, distill the patterns of neural network work recorded by the experiment, be inspired by them, infer how brain learning and thinking work, and generalize to the field of machine learning.
Sohu Technology: What do you think should focus on in the next stage of AI development? What are the shortcomings of current AI?
Wu Si: I think there are many things that need to be improved, if the goal of artificial intelligence is to build a human-like agent, then it also needs to have the ability to move, interact, execute, and adapt to environmental changes.
The technical route of imitating brain structure and realizing intelligence through neural networks is to be adhered to. In addition to the vision system, hearing, smell, perception systems, and motion control systems are all waiting to be realized by AI in the future.
There is a saying that there are about 400 cognitive functions in humans, and current AI has achieved less than 10, such as the function of processing and processing complex information in the subconscious mind. ChatGPT is just a language model, but not an agent, it is like a military master to give you ideas, but the execution still needs people to do, and the current AI has not yet achieved the so-called 'embodied intelligence'.
Another point is that the current AI model is almost all static, such as you give a command, it gives an output, but humans are not, our brain has a perception system to perceive external information, an emotional memory system to assist information processing, and a brain command to coordinate various behaviors, so AI still has a lot to do, what we need to construct is a truly like our brain, an independent, comprehensive information processing system.
Sohu Technology: A recent study found that employees' frequent use of AI at work may cause "potential physical and psychological harm" to employees, and researchers said that humans are social animals and that using AI systems to isolate work may have a devastating effect on employees' personal lives. What do you think of this view?
Wu Si: This obviously points to the serious social problems that AI may bring. When it comes to new technologies, such as cell phones, there's a hilarious picture of everyone sitting together eating and everyone looking at their phones but not talking to each other. Therefore, a new technology, when widely used, will bring a mixed impact to human society, but it must have advantages, otherwise no one will do AI.
The industrial revolution in history has completely changed human society, including social structure, marriage and family structure. AI may be the same in the future, its impact will be huge, and I hope that we will be a little prepared for future changes.