General Assembly | Yao Xin, Professor of Birmingham University: Three Forgotten Problems in Brain-like Computing Research

Lei Fengnet (search "Lei Feng Net" public concern) by: Yao Xin, is currently a lecture professor at the Computer College of the University of Birmingham, UK, director of the Center for Computational Intelligence and Applied Excellence, IEEE Fellow (Academician), IEEE Outstanding Professor of Computational Intelligence. . From 2003 to 2008, he was the editor-in-chief of the IEEE Evolutionary Computation Journal. From 2014 to 2015, he was the chairman of the IEEE Evolutionary Computing Society. His main research areas include evolutionary computing and integrative learning and its applications, especially in software engineering. His dissertation was awarded the IEEE Donald G. Fink Award for Outstanding Paper Award in 2001, the IEEE Evolution Computing Computational Distinguished Paper Award for 2010 and 2015, the BT Gordon Radley Best Creative Author Award for 2010, and the 2011 IEEE Neural Network Distinguished Paper Award, And some other best paper awards. In 2012 he was awarded the prestigious Royal Society Wolfson Research Award for Merit, and in 2013 he was awarded the IEEE Intelligent Computing Society Pioneer Award.

Professor Yao Xin reports on the 2016 Artificial Intelligence Hunan Forum

In recent years, brain-like computing has once again attracted the attention of researchers and the media. There have been many suggestions for building artificial brains, studying brain-like calculations, and some grand plans to understand how the human brain works. But what are the important scientific issues in brain-like research? Prof. Yao Xin discusses three issues that seem to have been overlooked by many types of brain computing researchers.

The first one is about evolution. All biological brains are evolutionary, but current brain-like computational research programs rarely involve the role of evolution. Should evolution be considered in brain-like computational studies?

The second is about the operating environment of brain-like computing systems. What is the role of the environment in brain-like computational research?

The third is about the role of the body or brain-body interaction. There is no brain without a body in the biological world. The current brain research program rarely mentions the role of the body, as if there is nothing to do with the body's brain research. Is this really the case?

Professor Yao Xin did not provide exact answers to the above questions. He only modestly expressed that he hopes to play a role in attracting these three issues.

No answer, first ask questions

These three issues are not so closely related to the industry, but they must also be taken into account when doing related industries. This background is very simple. At the opening ceremony, people may have felt that artificial intelligence is now quite prosperous. Not only is it in the research community and schools are prosperous, but it is also very prosperous in the industry. Leaders also said that Changsha City, Hunan Province, including Yuhua District, all like to introduce artificial intelligence-related industries, preferably smart robots, rather than general industrial robots.

From a research perspective, artificial intelligence is linked to the study of the human brain. According to this plan, the European Union has a special human brain research program; the United States also has a similar human brain research plan, and it is not called the brain. It is called the brain, including the human brain and the machine brain. It was also announced by President Obama; in the IEEE There is also a brain research program in the association, which is to promote international cooperation; there are also many domestic, including the Academy of Sciences or the national level want to study this brain.

Then, how to understand the brain is one aspect. On the other hand, it is how to truly apply the understanding of the brain to the project. As an artificial intelligence advancing role, what I am talking about today is not to say what answers are to be found in brain research or artificial intelligence research. I have no answer, but I have problems and there are not many problems. There are three problems. The questions are related to my research background.

The first question - creation or evolution

When people talk about artificial intelligence or the human brain, they always like to say that they have done an artificial intelligence, but when I do engineering or science people think about it, all the brains are evolved, and none of them are artificial. from. Now we want to create a brain. Of course, there is no problem. The only problem is a little philosophical. For example, now we want to create an artificial intelligence system that depends on people. Now we have to do a brain plan and want to be an artificial brain. At the same time, we need to find inspiring things from the natural brain. The natural brain is What evolved was not created. So the logic here seems to be rather strange. Although interested in a product, it ignores how the product came. This method is not right. If you are only interested in the product, no matter what the product is, only the products developed in the future will be studied.

In the study of neural networks, there is a very simple neural network called a hierarchical network, which is a layer by layer network. The simplest three-layer network, one input network, the middle is called the lead layer, and one is the output layer. This network can be as clever as all scientists, each input is either 0 or 1, and network learning can determine whether the 0101 string is even or odd. This is a very simple thing to say, but if you can only give an example to the computer, by learning, the future input will be 1 if it is an even number, and 0 if it is an odd number. Since this problem is more difficult, researchers want to use design artificial neural networks to accomplish this task. Indeed, this is a neural network designed by people. It is a very regular rule. And it's very easy to understand because there are only three floors. There are eight inputs, and then the nodes of eight neural networks are also designed. The output is 0101. So people design something very well structured and well understood. However, it is assumed that instead of designing a neural network, artificial evolution is used. Let it evolve an artificial neural network. What are the results and what are the results of artificially designed ones? The result of this evolution is nine inputs, not eight inputs, which are not the same as the previous ones because the hierarchy is not particularly clear and there is no direct connection.

Through the evolution calculation, it is found that the structure of a neural network and the network structure calculated by humans are actually very different and can be summarized as follows:

The first point is that the evolutionary computational network is very compact, that is, this input is not a real neural source. The true neural source uses four. The number of neural sources in the middle design is always the same as the number I input. So this neural network is particularly large. This is a problem that has really emerged from evolution.

The second thing is through the automatic evolution of the discovered neural network, the program will be a little more, not as regular as the artificial, except for the middle layer there is no.

The third point is that after this structure came out, it was a bit disorganized, so the left and right sides were asymmetric and not well understood. The problem here was more interesting. The so-called artificial neural network and the artificially evolved artificial neural network were designed by people. All solve the same problems, but the structure is not the same. This brings up a new problem, assuming that it is naive to regard it as the so-called small brain. This small brain has completed the measurement problem, but there is no structure here, where is the contradiction? This is very strange. It is interesting to do research here. How to explore this issue? It can be seen that what is now called artificial intelligence or brain-like computing or artificial neural network can play chess or go to the next world and take the first place in the world. Two days later, we saw that artificial intelligence can recognize the image manually and millions of images can be found. These things are different from what people have to do. A brain is responsible for all things, and all current neural network systems only focus on one thing. Alpha Go will only play chess and will not recognize the image. This is a very wonderful phenomenon. Everyone says to be smart, but the inspiration from people here has changed into the world of artificial intelligence. In today's artificial intelligence systems or artificial neural networks, a system does one thing, doing very well and very specifically. But a brain needs to do many things, so if a system does more than one thing, what will be the impact of the structure? This is related to what the brain module does. The static environment or the dynamic environment is not the same. This leads me to the second question I want to talk about.

The second question - artificial intelligence or artificial intelligence system

Although many researchers are interested in artificial intelligence or brain-like computing, they rarely consider what this artificial intelligence system can do. Light tells artificial intelligence, does not say what the artificial intelligence system can do, is missing something. There is still a difference between artificial intelligence and artificial intelligence systems.

One of the details that we wanted to talk about was the same as the first question. At least in the school, we did a small experiment first. To set up an artificial network system, let it learn one thing, which can be image recognition or other. Just let it do this, and then let the same neural network system learn two things at the same time, then observe the same initial state, and finally learn how different the structure of the neural network, there will be very interesting findings. This experiment can be repeated. After a neural network has completed multiple tasks, the modular structure will be clearly displayed. Regardless of the criteria used to measure the module, if you let a neural network do a job, some of the module functions are not fully functional. This means that when constructing an artificial intelligence system, you cannot talk about the system. You must talk about what the system needs to do, and do one or two things, whether in a static environment or in a dynamic environment.

The third problem - the carrier of artificial intelligence

This is actually more simple and it is a matter of the body. This problem is also very special, and all those who talk about artificial intelligence often do not tell where the artificial intelligence system was placed last. But all the brains are on the body, so the body is actually very important. During the study, the limbs and the six limbs affected the brain. Why research artificial intelligence often only studies the human brain and does not study the body, mainly because we really need a particularly developed brain when doing artificial intelligence research.

We have done an artificial experiment that artificially creates a nematode that can swim. This nematode is a section, and then the structure of each section is very simple. Each small circle represents a neural source. Some of this nerve source is used to control exercise. Muscles can contract and stretch. Then how does the nematode move? It can be imagined that the nematode travels like a wave type, that is, muscle contraction on one side and muscle contraction on the other side. After a period of time, the area where the contraction took place stretched, the area where it was stretched, and it slowly moved forward. Then let this artificial system swim straight from the right to the left. The faster you swim, the better, but it is not designed from the perspective of human thinking. It is to put the nematodes in the water to swim by themselves. There is a feedback for each tour. , record how many centimeters swim inside the unit time. What I want to observe is how this neural network will appear when I give this nematode different tasks and different postures, so I assigned two tasks to this nematode.

The first task is to let this nematode travel from A to B along a straight line, the faster the better. When designing a neural network controller, it is better to make this nematode swim as fast as possible, and the actual controller is very simple. Several small circles drawn correspond to the source of the nerve. The position of the nerve source can be adjusted. The position of the nerve source is different, and the strength of each contraction and stretching is different. We study what the structure of nerves is. This is just a section of it. The upper right corner looks like a small number, 0, 10, 200, 300 to 1190. This is how many generations I expressed using this algorithm to evolve artificial neural networks. The 0th generation represents initialization, initialization does not know how to design, so the location of all neural sources are randomly placed on this map. By the 10th generation, the structure began to appear a bit, because it was discovered that the connection of the neural source developed two broad, the left neural source was also connected, and the right neural source was also connected. At the 30th, 200th, and 300th generations, you discovered that there was a certain pattern. There was no direct connection between the nerve sources, the left side and the right side. This was not something people came up with. It was developed by people. By the 1190's, unconventional whole-symmetry structures came out. That is, the small distances on the two sides are almost the same. This is not something people have designed but discovered. from. So this is quite interesting. At first it did not give any indication to the evolutionary algorithm, but given the nematode structure, the neural network that emerged was a very symmetrical neural network.

The second task is to study the relationship between neural networks and posture in this case. The worm's posture is limited, but the task still has to go straight up, from A to B. Everyone can imagine that it is impossible to do this kind of posture under normal circumstances. Because I go to the right in this way, I will not go forward. But when I was experimenting, I didn't tell it about this algorithm, just told it that it was a place where the body structure is now limited. When the body is going to squat, the short side shrinks and the long side stretches, certainly not symmetrical. If it is symmetrical, it will not move forward in the direction of the head. This is a very small example, indicating that designing an artificial neural network or general exploration of intelligent systems is actually closely related to the physical posture.

The most fundamental point is that when researching artificial neural networks, it is necessary to consider what kind of physical system the final neural network is placed into. For example, if you want to study the robot itself, and at the same time design a system that controls the robot, then the posture and control of the robot are closely related and cannot be considered separately.

to sum up

The three issues are summed up in three sentences:

One is that the brains of all living creatures have evolved and are not created by God. When designing artificial brains, should we consider more about the process of evolution and learn something from the evolutionary process?

Second, at least in the natural world, the human brain must be able to perform multiple tasks, and this is done in a dynamic environment or an uncertain environment. Many of the artificial intelligence systems that are being done now define the definition on a very narrow and very specific function, such as recognizing images or playing chess. These two are actually artificial intelligence systems that are actually designed for the future. It has a very different influence. This issue deserves our careful consideration.

Third, all the brains have a carrier in the biological world. That is the body. There is no light, brain, or body. This poses a new challenge for us to build intelligence in the future. We must consider the carrier when designing artificial intelligence systems.