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

Https://blog.csdn.net/cf2SudS8x8F0v/article/details/90986936

Source: economic Observer

Recently, Professor Zhang cymbal, academician of the Chinese Academy of Sciences and director of the artificial Intelligence Research Institute of Tsinghua University, said in an interview with reporters that the current artificial intelligence based on deep learning has reached the ceiling. In the long run, we must take the road of human intelligence, and eventually develop a world in which human and machine coexist harmoniously. In the future, it is necessary to establish interpretable and robust artificial intelligence theories and methods, and to develop safe, reliable and credible artificial intelligence technology.

Academician Zhang cymbal: the miracle of AI is difficult to reproduce in the short term. The potential of deep learning technology is nearing the ceiling.

Three years after Alphago won the game against South Korean go player Lee se-dol, some signs gradually emerged, and academician Zhang cymbal thought it was the right time and accepted this exclusive interview.

At present, deep learning is the most concerned area of artificial intelligence, but it is not the whole research of artificial intelligence. Zhang cymbal believes that although there is still room at the industrial level, the current deep learning-based artificial intelligence has reached the ceiling in technology, and the "miracle" brought by this technological route did not appear again after the victory of Alphago, and it is estimated that it will be difficult to continue to appear in large numbers in the future. Technological improvement is difficult to completely solve the fundamental defects of artificial intelligence in the current stage, and these defects determine that the space for its application is limited to specific areas-most of which are concentrated in image recognition and speech recognition.

At the same time, in Zhang cymbal's view, at present, the business community and some academic circles all over the world are too optimistic about deep learning technology, and artificial intelligence urgently needs to be promoted to a new stage, which is destined to be a long process. It depends on the breakthrough of the underlying theory in combination with mathematics and brain science.

As a rare researcher in China who has experienced two stages of artificial intelligence technology, Zhang cymbal has given few interviews in the past few years, one of the reasons is that he has some different views on the current development of artificial intelligence technology. before the time has come, Zhang cymbal cautiously believes that these views are not easy to spread through the mass media, and even if spread is difficult to be recognized.

1. "miracles have not happened, and according to my estimation, they will not continue to occur in large numbers."

Economic Observer: how do you estimate and evaluate the current development of artificial intelligence?

Zhang cymbal: this upsurge of artificial intelligence sprang up at the beginning of this century. First of all, it appears in the academic circles. In the past, the academic circle was indifferent to artificial intelligence, but the emergence of multi-layer neural network has brought some changes. The theory of neural network has been available since the 1950s, but it has been in a state of shallow application. People did not expect that multi-layers would bring any new changes.

What really attracted everyone's attention was the Stanford experiment in 2012. In the past, the number of image samples was at most "ten thousand". Stanford used 10 million, using multi-layer neural networks to do it, and it was found that in the three image categories of human face, human body and cat face, The recognition rate of this model has been improved by about 7% and 10%.

This is a great shock to everyone, because usually it takes a lot of efforts to increase the recognition rate by 1%, but now only the number of layers has been increased, and two major changes have taken place: one is that the recognition rate has been improved so much; the second is to be able to deal with such big data. These two changes are very encouraging, not to mention that artificial intelligence did not solve practical problems before 2012.

Economic Observer: what is the reason for this breakthrough?

Zhang cymbal: now there are three reasons, and everyone is very clear: one is big data, one is computing power, and the other is algorithm. After realizing this, overnight, both inside and outside the industry were very shocked by deep learning, and then three historic events took place.

The first thing is that in December 2015, Microsoft reduced the image recognition error rate to 3.57% through a 152-layer deep network, which is lower than the human error rate of 5.1%; second, the word error rate of speech recognition done by Microsoft in 2016 was 5.9%, the same as that of professional stenographers; and the third thing: Alphago beat South Korean go player Lee se-dol.

Through artificial intelligence, using the two tools of deep learning and big data, we can surpass human beings under certain conditions and in certain fields. These three things give us great encouragement.

Especially for people outside the industry, they all think that as long as I have mastered big data, I may be able to work miracles through deep learning, so we have made a lot of predictions, such as how soon computers will surpass people in what things.

But in fact, after this, miracles did not happen, and according to my estimation, there will not be a large number of them in the future. To be precise, progress may be made in individual areas in the future, but it will not blossom as fully as previously expected. In particular, the Chinese market is optimistic that "the Chinese market is large, there are many data, and the use is unrestricted, so miracles will happen in China in the future."

As a result, many enterprises find that this is not the case when they do it. From the current situation, the best things are these two things: image recognition and speech recognition. I took a look, 20 unicorns and 30 quasi-unicorns in the field of artificial intelligence in China, nearly 80% of them are related to image recognition or speech recognition.

Economic Observer: why did this happen? Or after such a long time, do we have a clear understanding of what artificial intelligence can do now?

Zhang cymbal: this panic arises after artificial intelligence defeats humans in go. "what a master can do, artificial intelligence can do. My work is so ordinary that it will certainly be replaced by machines." Its limitations need to be considered here, and I have been talking about not being too optimistic at various meetings.

The three things that artificial intelligence can do (speech recognition, image recognition, go) is because it meets five conditions, that is, as long as these five conditions are met, the computer can do it well, as long as any one or more conditions are not met. It's difficult for the computer to do.

The first is that we must have sufficient data, which means not only a large number of data, but also diversity, not imperfection, and so on.

The second is certainty.

The third is the most important, which requires complete information. Go is a game with complete information, and cards are games with incomplete information. Go is complex, but in essence, it only needs fast computing and does not rely on any intelligence, but in daily life, all our decisions are made under incomplete information.

The fourth is static, including the law of evolution according to certainty, which is the problem of predictability, which is not satisfied by autopilot in complex road conditions; in fact, it does not satisfy neither certainty nor complete information.

The fifth is a specific field, which he cannot do if the field is too wide. A single task, that is, artificial intelligence software for playing chess, is playing chess and can do nothing else.

Economic Observer: that is to say, under the premise of meeting these five conditions, the current artificial intelligence is competent for part of the job?

Zhang cymbal: if your work meets these five conditions, it will definitely be replaced by a computer. the working characteristics that meet these five conditions are very obvious, that is, the four words "do things by the rules" and do not require flexibility, such as cashiers and cashiers. If your work is flexible and creative, the computer can never be completely replaced, of course, partial replacement is possible, because there must be some simple and repetitive content. If you recognize this principle, you will realize that artificial intelligence is still in the early stages of development. Not as some people estimate that "artificial intelligence technology has been fully mature, and entered the stage of development and application."

2. "Deep learning technology is close to the ceiling from the point of view of application."

Economic Observer: how should we define the current deep learning technology route? is it a thing based on probability?

Zhang cymbal: the essence of deep learning now is based on probability and statistics. What is called probability and statistics? It is not so mysterious. Deep learning is to look for patterns that occur repeatedly, so if you repeat more, it is considered a law (truth), so a lie repeated a thousand times is considered a truth, so why does big data sometimes make very ridiculous results? because no matter whether it is right or not, as long as it is repeated more, it will follow this law, that is, whoever says too much is the one who says too much.

I often say that we have not yet entered the core issue of artificial intelligence, in fact, the core of artificial intelligence is knowledge representation, uncertain reasoning, because where is the source of human intelligence? In knowledge, experience and reasoning ability, this is the foundation of human reason. The artificial intelligence systems formed now are very fragile and vulnerable to attack or deception, need a lot of data, and can not be explained, there are very serious defects, this defect is essential, caused by the method itself.

Economic Observer: that is to say, it cannot be solved completely through improvement? For example, if we increase the number of layers and complexity of the neural network or increase the order of magnitude of the data, will its defects be solved?

Zhang cymbal: improvement is not good. The essence of deep learning is to use the "black box" processing method of probability learning to find its rules by using unprocessed data. This method itself usually cannot find "meaningful" rules. It can only find recurring patterns, that is to say, you cannot achieve real intelligence by relying on data alone.

In addition, deep learning is only a part of the current artificial intelligence technology, and artificial intelligence has a larger and wider field to study, such as knowledge representation, uncertainty processing, human-computer interaction, and so on. We can't say that deep learning is artificial intelligence, deep learning is only a part of artificial intelligence. Until last year, 1/3 of the papers exchanged at the artificial intelligence conference were on machine learning and 2/3 on other aspects.

Economic Observer: do the academic circles still have a relatively clear understanding of this?

Zhang cymbal: I can say that most of the academic circles around the world have a clear understanding; most of the business community around the world are overly optimistic.

Why did this happen? Because most of the people who have been engaged in early artificial intelligence research are dead or old and have no right to speak. Now active in the front line of artificial intelligence research are deep learning, big data joined after the rise, their understanding of artificial intelligence is not comprehensive enough.

Economic Observer: if every technological route has a "technological potential", how much have we used this potential in terms of deep learning?

Zhang cymbal: scientific research is difficult to estimate accurately, but if deep learning does not change it from an application point of view, I think it is close to the ceiling, that is to say, you are less likely to have another miracle.

Economic Observer: based on this, do commercial companies still have a lot of room for underlying technology and industrial applications?

Zhang cymbal: as long as we choose the appropriate application scene and use mature artificial intelligence technology to do the application, there is still more room for application. At present, in the academic circles, in order to overcome the problems of deep learning, we are carrying out in-depth research work, hoping that the business community, especially small and medium-sized enterprises, will pay close attention to the progress of the research work and apply new technologies to their own products in a timely manner. Of course, enterprises of the size of Google and BAT will all engage in related research work, and they will combine research, development and application.

Economic Observer: there is a view that the "white box" (understandability) we emphasize is actually emphasized from the human mind, but through big data, probability and statistical tools discrete to continuous projection, it is actually the thinking of a machine. You don't necessarily need it to give you an explanation, just the right answer?

Zhang cymbal: at present, there are two opinions, one is that there are many ways to be intelligent, not only one way can lead to intelligence, we have produced natural intelligence through natural evolution, so why can't we generate machine intelligence through machines? This intelligence and natural intelligence will not be exactly the same, all roads lead to Rome, and the intelligence we gain through natural evolution is not necessarily the best. I agree with this point of view. Machine intelligence is different from human beings. In fact, it is beneficial to complement each other and give full play to their respective strengths.

But in the long run, we have to take the road of human intelligence. Why? Because in the end, we want to develop a world in which man-machine cooperation and human beings and machines live in harmony. We are not saying that machines will take care of everything in the future, and human beings will enjoy it. We have to take the road of man-machine symbiosis, so that the intelligence of machines must be the same as that of human beings, otherwise they cannot live together. We can't understand what machines do, and machines don't know our intentions. How can the two cooperate?

Economic Observer: it must be explainable?

Zhang cymbal: yes, it is explainable. You want it to make a decision. If you don't understand it, the plane will fly. Who dares to take this plane? So at the present stage, cars and airplanes cannot be completely driven by machines. Why do we rest assured that the driver sits on it? Because we share the same fate with him, we are going to be killed together. The machine is not the same fate as you. It can't be killed. You killed it.

Some people are very divorced from reality to think about this problem, this is not right, how can human beings develop machines like that? Human beings will not develop like that. Some people are worried about what robots dominate human beings. I say this can only be regarded as far-sighted at best.

Economic Observer: so Turing's paper also said that this view is "not worth refuting."

Zhang cymbal: yes, that is a long-term concern. At present, we still have a lot of immediate worries. Security must be considered in the development of artificial intelligence, which is already a realistic problem.

If you look at speech synthesis, you can make use of the existing technology to make it look real, which is basically the same as that of a real person. Now it seems that this technology can not be popularized and applied, because once promoted, it will be a complete mess, as long as a fake recording made of speech synthesis technology can discredit any celebrity. These are very dangerous technologies. The governance of artificial intelligence is already on the agenda.

Third, "We cannot cultivate Einstein or Turing."

Economic Observer: one view is that China has more data and more engineers. Can this scale lead to breakthroughs in basic research or determine the route of technology?

Zhang cymbal: there are a lot of concepts confused here, science, technology and engineering. The scientific and technological level needs to be measured by three standards, one is the scientific research level, the other is the technical level, the other is the engineering practice ability, or the industrialization ability.

What is the situation in China? From an engineering point of view, we are "close to the world level" in some areas; the word I use in technical level is "large gap", because many things can still be done by foreign countries and we can't do them. The word I use in the field of scientific research is "a big gap". Scientific research is original. In fact, all the original achievements in the field of artificial intelligence are made by Americans. There are 11 Turing Award winners in the field of artificial intelligence, ten Americans and one Canadian.

Economic Observer: the data show that the number of papers published and the number of citations in the field of artificial intelligence in China have entered the forefront. Does this indicate a breakthrough in the field of artificial intelligence in China?

Zhang cymbal: if you look at the research level of the paper alone, it is basically reflected in three indicators: quantity, average citation rate, and the highest citation rate of a single article. In the case of artificial intelligence, the number and average citation rate of Chinese researchers' papers are good, but there is a big gap between the highest citation rate of a single article and the world, and this indicator precisely reflects your original ability.

In other words, in the field of deep learning, our average level has reached the world level, but there is still a big gap between the highest level and the world. However, we have to be sure that we are developing relatively fast in application.

Economic Observer: does Tsinghua have any advantages in this respect?

Zhang cymbal: in the important conference journal of artificial intelligence, during this decade, the number and average quality of papers CMU (Carnegie Mellon University in the United States) ranked first, and Tsinghua University ranked second. The people we train, in the field of computer, the undergraduate and doctoral students of Tsinghua University are world-class.

At present, our tracking ability is relatively strong, once someone starts, we can quickly follow. But unfortunately, we lack top people, and we can't train top talents, such as Einstein, Turing and so on.

I personally think that one of the reasons may have something to do with Chinese culture. our herd mentality is very serious. for example, in the field of artificial intelligence, deep learning is very hot, and almost 70% of the published papers are Chinese, but there are few Chinese authors in other non-hot areas, including uncertain reasoning and knowledge representation. This is a crowd gathering, unwilling to explore the "no man's land".

Of course, don't worry. Scientific research is originally done by the rich and by rich countries. We are still developing countries, and since the starting point of scientific research is relatively low, it is inevitable that we will lag behind temporarily, and we will catch up.

4. "the low ebb will happen, but it will not be like in the past."

Economic Observer: if deep learning has reached the ceiling, where will the future direction of artificial intelligence be?

Zhang cymbal: recently we are going to put forward a new concept, that is, the concept of the third generation of artificial intelligence. Artificial intelligence has actually experienced two generations, the first generation is symbolic reasoning, and the second generation is the current probability learning (or deep learning). We think we are entering the third generation of artificial intelligence. The reason is obvious that both the first generation and the second generation have great limitations.

Economic Observer: does the third generation artificial intelligence technology you are talking about have a clear direction or characteristics?

Zhang cymbal: what we propose now is to establish interpretable and robust theories and methods of artificial intelligence and to develop safe, reliable and credible artificial intelligence technology.

Economic Observer: such technology may have to wait a long time?

Zhang cymbal: yes, it's hard to predict. We're in a hurry, too.

Economic Observer: do we have to return to mathematics and other theoretical levels to find new methods?

Zhang cymbal: at present, we have two ways, one is to combine with mathematics, the other is to combine with brain science. If you think about it, if there are no new mathematical tools, no new ideas inspired by brain science, how can there be new theories? On the other hand, it is necessary to combine data-driven and knowledge-driven, because it is difficult to find a breakthrough in mathematics and brain science.

Economic Observer: does this combination refer to the integration of previous decades of artificial intelligence experience?

Zhang cymbal: yes, at least one direction is to combine the first generation with the second generation and make use of their respective advantages. But the combination of the two is very difficult, because they operate in different spaces, one is the vector space, the other is the symbol space, and new mathematical tools are needed.

Economic Observer: look at the history of artificial intelligence, there is a long interval between each generation of technology, will the third generation of artificial intelligence technology be the same?

Zhang cymbal: I think it will be longer, because it needs to be tackled, because the problems encountered are more difficult.

Economic Observer: in another 10 or 20 years, will artificial intelligence become a "hidden science" in the minds of the academic community or the public, just like in the 1970s and 1980s, and the public will no longer mention this word often?

Zhang cymbal: the low ebb will happen, but it will not be like in the past. What is the reason? Because of big data, the Internet and powerful computing resources, all these will support artificial intelligence to continue, although sometimes it is only superficial prosperity.

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.