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

Https://www.toutiao.com/a6705912318187799052/

China artificial Intelligence Conference (Chinese Congress on Artificial Intelligence 2019, referred to as "CCAI 2019") will be held in Jiaozhou, Qingdao, from September 21 to 22.

Professor Zhou Zhihua is one of the chairmen of the procedural Committee of the Congress. He has been engaged in artificial intelligence research for more than 20 years, including artificial intelligence, machine learning and data mining, and is the representative of artificial intelligence research in China. Professor Zhou Zhihua is currently the dean of the School of artificial Intelligence of Nanjing University, the first Chinese "Grand Slam" Fellow of the five major international societies related to artificial intelligence, and a foreign academician of the European Academy of Sciences.

Artificial intelligence represented by deep learning has made remarkable development in recent years. As a world-class expert in this field, what kind of observation and thinking does Zhou Zhihua have, and what kind of disciplinary insights has he put forward? Let's take a look.

Three Secrets of success in Deep Learning

In recent years, the popularity of artificial intelligence is inseparable from several breakthroughs in deep learning, especially the emergence of go intelligence AlphaGo. With an overwhelming advantage, it has swept the two top human chess players, Li Shishi and Ke Jie, and while gaining a lot of surprise and admiration, it has aroused people's infinite imagination of the ability boundary of artificial intelligence and the possibilities of life in the future.

As a world-class expert in the field of machine learning, Zhou Zhihua has always maintained a highly rational and sober understanding of the development of technology. He attributes the current success of machine learning to these three factors: an effective depth model, strong supervisory information and a stable learning environment.

At the present stage, the effective depth model is basically the depth neural network. It has been pointed out that neural networks are not new, they existed as early as half a century ago, and today we can do deeper neural networks only because of our strong computing power and can now be trained.

Zhou Zhihua believes that this is a misunderstanding. He pointed out that before 2006, scholars did not know how to train neural networks with more than five layers, not because of computing power constraints, but because they did not know how to design effective algorithms. Turing Award winner Jeff Hinton made an important contribution in this regard, making it possible to train deeper neural networks, which led to the development and prosperity of deep learning.

The second factor is the existence of strong supervisory information. Deep learning requires a large number of samples, so in today's big data era, samples are not a problem?

Zhou Zhihua explained that it was not. Only samples are not enough, the important thing is that these samples need to be marked, which will cost a lot of manpower and material resources, and machine learning technology is highly dependent on strong supervisory information.

Take AlphaGo as an example, it uses more than 160000 chess games of more than six segments of the human profession for training. AlphaZero, which was later invented, does not use human chess games and uses two programs to directly improve its performance, but this model also depends on strong supervisory information, because it depends on the rules for judging the outcome provided by human beings, and this rule itself is very strong supervisory information that is difficult to have in general application tasks.

A stable learning environment is also indispensable to the success of current machine learning technology. In such an environment, the data distribution, sample space and learning goals involved in the learning process are all fixed.

How to achieve a breakthrough in deep learning?

In Zhou Zhihua's view, any model is bound to have defects, and so is the deep neural network. First, it requires people to spend a lot of energy to adjust the parameters; second, the neural network learning model used now needs to be determined in advance before training, but in fact, people do not know what kind of complexity model is the most appropriate before solving a practical problem. In addition, the dependence on large training data, the difficulty of theoretical analysis, the black box of the model and so on can not be ignored.

Zhou Zhihua pointed out that "there is no free lunch theorem" in the field of machine learning shows that there is no model of "dominating the world", and that only part of the tasks of any model may be applicable. It can be seen that the depth neural network has excellent performance in image, sound and other numerical modeling tasks, but it is not outstanding in other tasks such as symbolic modeling, discrete modeling, hybrid modeling and so on. it is necessary to consider the design of a new deep learning model besides the neural network model according to the characteristics of such tasks.

Secondly, the current deep learning is highly dependent on strong supervisory information, but strong supervised samples in real life need to pay a huge price, how to use weak supervised information for learning is an important research direction.

Finally, more and more people come across an open and dynamic environment, and the factors involved in the learning process may change. This poses a greater challenge to machine learning.

Therefore, Zhou Zhihua concluded that more attention should be paid to machine learning research in the future:

A new deep learning model outside the neural network; learning based on weak supervised information; learning in an open dynamic environment.

Talent training of artificial Intelligence

Zhou Zhihua said frankly that in terms of its historical contribution to artificial intelligence, China is not only inferior to the United States, but also inferior to Europe and Japan. In the world, artificial intelligence research began in the 1950s and 1960s, while China resumed scientific research after 1978, which started too late, and artificial intelligence research began to be in line with international standards since the new century.

At the same time, he said that China has made rapid development in artificial intelligence in the past decade, and if we only look at the development in recent years, it can be said that China is second only to the United States in the field of artificial intelligence.

In 2018, the first artificial intelligence college of C9 universities in China was established in Nanjing University, with Zhou Zhihua as the first dean. By March 2019, 35 universities had obtained the qualifications for the first batch of undergraduates majoring in artificial intelligence.

Zhou Zhihua believes that the emergence of artificial intelligence colleges is the result of policy promotion, industrial development and discipline connotation: first, artificial intelligence is the focus of the times, and the national government attaches great importance to it; second, artificial intelligence technology has entered the Internet, finance, intelligent manufacturing and many other industries, the talent gap is very large. Third, artificial intelligence is not a short-term hot spot, but a serious discipline after more than 60 years of development, has formed a huge self-consistent professional knowledge system, a comprehensive grasp of artificial intelligence professional knowledge requires a lot of learning time and energy.

In terms of talent development, Zhou Zhihua pointed out that artificial intelligence is an industry that highly highlights "personal heroism". Because in the past, for talent training, the gap between "learning" and "using" existed naturally. Take software development as an example, almost no software can be completed by only one or two people, and it must be done through large teams, while schools mainly cultivate students' personal quality and ability. for example, it is impossible for dozens of students to do the same test paper and share scores in a course examination, and students' ability to work in a large team must be cultivated and developed in an enterprise. The artificial intelligence industry is different, its "learning for application" is very strong, high-level talents trained from colleges and universities can quickly play a role in the industry. Zhou Zhihua led LAMDA (Institute of Machine Learning and data Mining of Nanjing University) to train graduates to join Ali, Tencent and Huawei to make important breakthroughs for the company. A number of doctoral students were even employed as directors of well-known enterprise research institutes during their studies, which is an obvious example. Zhou Zhihua pointed out that the "path" in the field of artificial intelligence from laboratory results to industrial applications is quite short, and the algorithm breakthroughs made by individual "smart people" in the laboratory can often be quickly implemented in industrial applications, and even promote and lead the technological development of the whole field. In order to speed up the development of science and technology and industry in China in the era of artificial intelligence, colleges and universities can make greater contributions in training high-level artificial intelligence talents.

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.