Shulou(Shulou.com)06/02 Report--

2019-11-10 04:31:39

Author | Camel

Editor | Tang Li

On October 31, the first Zhiyuan Conference was held by Beijing Zhiyuan Institute of artificial Intelligence, which was positioned as a "AI event for experts". At the meeting, many famous scholars in the field of artificial intelligence, such as Academician Zhang cymbal, Academician Gao Wen, John E. Hopcroft, Michael I. Jordan, Zhu Songchun, Christopher Manning and so on, provided a big stage for the cutting-edge exchange of domestic scholars in the field of artificial intelligence.

At the meeting, Academician Zhang Pingwen of the Mathematics Department of Peking University, as the chief scientist in the major research direction of "the Mathematical basis of artificial Intelligence" of Zhiyuan Research Institute, was interviewed by the AI Science and Technology Review and other media.

Zhang Pingwen, academician of the Chinese Academy of Sciences, academician of the Academy of Sciences of developing countries. Professor and doctoral supervisor of the School of Mathematical Sciences of Peking University, special professor of the Yangtze River of the Ministry of Education, national candidate for the tens of millions of talents project, academic leader of the innovation group of the National Natural Science Foundation of China. He is currently assistant to the president of Peking University, director of the Science and Engineering Computing Center, director of the Chinese Society of Industrial and Applied Mathematics, director of big data National Engineering Laboratory of Analytical and Applied Technology, and director of the academic Committee of Beijing big data Research Institute.

During the interview, Academician Zhang Pingwen pointed out that "the mathematical basis of artificial intelligence" is the first research direction established by Beijing Zhiyuan artificial Intelligence Research Institute. He mentioned that the current artificial intelligence is facing theoretical challenges such as calculability, interpretability, generalization and stability, and the solution of these problems needs to be deeply considered from the perspective of mathematics, statistics and calculation. therefore, the major research direction of "Mathematical Foundation" will first start the research in three aspects, including interpretable new artificial intelligence models, new machine learning algorithms and the basic theory of deep learning.

At present, the progress of artificial intelligence often comes from researchers in computer or neuroscience. Academician Zhang Pingwen believes that Zhiyuan scholars will strive to break this paradigm and establish a new generation of artificial intelligence methodology based on mathematics and statistical theory.

In the subsequent interview, the reporter and Academician Zhang Pingwen also talked about some problems about personnel training. Academician Zhang Pingwen believes that at present, China has a leading position in the application field, but there are still deficiencies in basic research, mainly due to the lack of innovation environment.

As one of the best mathematics mecca in the country, the mathematics department of Peking University is crowned as "the first of the four madhouses of Peking University". Talking about the cultivation of talents in the mathematics department of Peking University, Academician Zhang Pingwen believes that there should be talent and feelings for the research of basic mathematics, both of which are indispensable, so the mathematics department of Peking University has never been to cultivate genius, but to protect it. Through screening to find all kinds of "geeks", protect them, and guide them to "don't do small problems, do big problems", "either blockbuster or obscurity". But big problems mean that most people are unsuccessful, so most people may be unknown all their lives, but Peking University has this spirit of "big problems" in its bones.

Let's take a look at the specific conversation of Academician Zhang Pingwen. The AI Science and Technology Review has been edited without changing its original intention, and the content has not been confirmed by me.

(participating media: AI Technology Review, CSDN, Caixin, 36 krypton)

Four levels

Zhang Pingwen: "Mathematical Foundation of artificial Intelligence" is the first research direction established by Beijing Zhiyuan Institute of artificial Intelligence. Its main purpose is to study the mathematical basis of artificial intelligence. We all know that there are still many problems in artificial intelligence, such as the interpretability of deep learning and so on. So, how can we really understand it? In fact, the basis of artificial intelligence is mainly mathematics, because the whole computer technology is mainly mathematics. But there are many levels of mathematics, and in general I divide it into four.

The first is the algorithm level. There are many algorithms in artificial intelligence, and deep learning is only one of them. This wave of artificial intelligence is relatively hot, mainly driven by deep learning algorithms. This algorithm works well, but it is not good for all problems. Secondly, we do not know why it is good. Those of us who do math hope to be able to figure out why this algorithm is good and what its scope of application is from a mathematical point of view. We also hope to be able to propose new algorithms based on mathematical research, but this problem is relatively difficult.

Another level is the model. In the past, mathematical models generally came from physics. Physics first found various mechanisms from experiments, and then expressed the mechanism in mathematics to become a mathematical model. We hope to explore a new mathematical model from artificial intelligence.

The second is the problem-driven level. Artificial intelligence now has a lot of problems that can not be solved. In the process of solving these problems, if the algorithm is not suitable and the model is not suitable, you will try to find a new algorithm, which we call problem-driven. In fact, this is exactly what a lot of computer people do these days, not math people. But people who do math should also be able to do a little, and if they don't do these jobs, it will be difficult for them to understand these problems.

However, none of the above is the highest level of mathematics, and the highest level of mathematics should return to basic mathematics. What is basic mathematics? As we say geometry, you have to look at the structure of the neural network from the perspective of geometry and topology, which is at a deeper level. But it is not clear which basic mathematics has something to do with the foundation of artificial intelligence, and all the connections are very weak. It is because it is not obvious that it is worthy of our in-depth study.

At present, people's understanding of what is the mathematical basis of artificial intelligence is still too far away. But on the whole, the mathematical basis of artificial intelligence is later than technology. This is not a unique phenomenon of artificial intelligence, all areas are like this, mathematics will not walk in front of technology, often the foundation walks behind technology. After we understand more about technology through mathematics, we will make the application more extensive and more confident in application.

Therefore, the research on the mathematical basis of artificial intelligence can not be regarded as an engineering. If the project is studied, it will not produce any good results; in fact, it is a very long process. For such research, it is more important to establish a good environment; in addition, such research is the world of young people, so we should attract particularly outstanding young people to work in this field and give them a good environment so that they can feel at ease to do this kind of problem. This is exactly what Beijing Zhiyuan artificial Intelligence Research Institute is going to do.

Basic innovation environment

Q: compared with some other countries (such as the United States, Canada, etc.), there is still a lack of AI research in China. What do you think of this phenomenon?

Zhang Pingwen: many people have talked about China's advantages and disadvantages in the report.

China's strengths, the first is data, the second is the application scenario. At the data level, China has a great advantage, because we have 1.4 billion people, and our privacy protection of data is not as high as that required by Western countries, so we have a great advantage in this respect. In terms of application, the application scenario in China should be very rich, coupled with some incentive policies of the Chinese government, the application is very strong.

The worst aspect in China is basic research, and the most typical is the original algorithm. The algorithm is equivalent to a bridge, you have a lot of data, but also have application scenarios, but how to achieve, this requires the algorithm of this bridge. But you can see that few of the important algorithms in artificial intelligence come from China, which will lead to a lack of understanding of the algorithms.

So in fact, both data and application, China is very good, compared with the United States is not bad, but the difference lies in the original.

Why is it bad in the original piece? I think there are many factors. On the one hand, Western countries (especially the United States) have attracted the best talents in the world, including Chinese talents, especially high-end talents, so that most of the world's outstanding talents are in the United States. Why can the United States attract these people? There are many factors, but the most important thing is that it has an innovative environment, the whole society is open enough, and there are the most senior scientists. If young students want to learn the most cutting-edge knowledge, they will give priority to the United States. Now 80% and 90% of the best math students in China are in Peking University, and half of these students will go to the United States, and they are the best half. If you want to keep them, you don't give them as much money as the United States, and they will stay. Excellent students must wonder whether I can learn the most cutting-edge things in a place and whether I have a good environment. This is the most important thing.

China is lacking in this respect. First, there is a lack of high-end people to teach them; second, there is a lack of an innovative and developing environment. Our country already has some, but it is not enough. What Beijing Zhiyuan Research Institute should do is: first, with the support of the scientific research environment, outstanding talents can work at ease in Beijing without being trapped by money. Second, through Zhiyuan, this platform enables people in different fields (you see, we do mathematical foundation, and others do chips, information retrieval, and natural language processing) to unite and learn from each other. This creates a good environment for innovation. I think these two points are very important. If Zhiyuan solves these two problems, we will be able to stabilize high-end talents and have senior scientists to educate our young people.

Of course, Zhiyuan can not solve all the problems, the training of talents ultimately depends on the school.

The growth of Mathematical Genius in Peking University

Not to cultivate, but to protect.

Q: so, how does the Mathematics Department of Peking University train talents?

Zhang Pingwen: in fact, the field of mathematics in China is different from other fields. now mathematics is extremely popular in Peking University. If you want to enter the science and engineering department of Peking University, the most difficult department is the Department of Mathematics. It has been like this for more than ten years.

From our own point of view, more than half of the students in our mathematics school are recruited through math competitions. we are not examining his mathematical ability in middle school, but his mathematical talent, because if a student wants to do basic mathematics in the future, the most important ability is the person's talent, basic math is not what anyone wants to do. In addition, there should be feelings, because the research of basic mathematics is driven by feelings and cannot be mixed with our oil, salt and vinegar.

So we have to find people who are talented and affectionate and want to do basic math all their lives. Some of these people are particularly good at basic math, some are similar to us, and some seem to be completely ignorant of the world. There are many "weirdos" in the mathematics department of Peking University. We have such an environment in Peking University, so that these "weirdos" can survive in Peking University, not to say that they want to survive well, but to be able to live a decent life and live a decent life.

Therefore, the first thing we have to do is to find such people, the second is to protect them, and the third is to give them an environment to grow up. You do not train these people. You want to create an environment in which they can grow. I think it is not accurate to use "training", because he grows up by talent and feelings.

There are too many temptations in the outside world, how to let these temptations not affect them, this is what we have to do. If he wants to learn something, he can learn it, and there are excellent scholars to communicate with them. But if he can't stand the temptation of society, watching the beautiful woman come and run with the beautiful woman, and when he sees the big boss coming, he will go with others to make money, won't it be finished? These people are very capable, we can not let them run away, we have to "protect" them.

These people are at our core. Ten or eight a year is enough, not too many, but we don't know who they are. We need to constantly identify them. This is our top priority. I will not say the details, because I do not want to make it very lively, their world needs to be very quiet.

We recruit almost 200 people a year, but as a matter of fact, 80% to 90% of them will have to enter society in the future. There are two popular choices, either artificial intelligence or finance. Don't these people need talent? Talent is still needed, but it's not that important. At this time, his more important ability is the ability to transform engineering problems into mathematical problems, the ability to solve mathematical problems, and the ability to apply existing mathematical knowledge to practical problems. Therefore, the training of them is not exactly the same as that of the previous minority.

Such people should first of all learn mathematics well, which can be cultivated. But the most important thing is to cultivate his mathematical thinking ability, can not completely turn him into a nerd, need to let them combine with the industry, the specific industry depends on his interest. In addition, he needs to have the ability of teamwork, basic math can fight alone, but applied mathematics needs a team.

Therefore, to train talents is to constantly screen, protect those essence and "geeks", and cannot publicize them. Our most talented people, we do not publicize, for fear of destroying his state of mind. You don't usually see these people. It's weird. You don't even know when you see them. In the future, there will be plenty of people like Zhang Yitang among them.

Peking University is full of spirit, we tell these people, you do not do small problems, do big problems; either blockbuster, or unknown, that's all. It's not easy. Making a big problem means you don't succeed a lot. As a result, most people may remain unknown for the rest of their lives. Peking University alumni are very proud, like the older generation Zhang Yitang, his experience can be made into a movie, 58-year-old finally became famous, the way is very bumpy.

There is this spirit in the heart of Peking University.

Https://www.toutiao.com/i6757407334038766083/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.