2065438+On July 1 day, 2009, the eighth research center of Tsinghua University Institute of Artificial Intelligence: Research Center for Natural Language Processing and Social Humanities Computing (hereinafter referred to as the Center) held its inauguration ceremony in the FIT building of the school headquarters. The director of the center is Professor Sun Maosong, executive vice president of Tsinghua University Institute of Artificial Intelligence and chief scientist of the National Key Basic Research and Development Program (973 Program).
On July 3rd, our reporter (www.thepaper.cn) walked into the campus of Tsinghua and chatted with Professor Sun Maosong about the background and mission of this exceptionally long center, as well as the development status, difficulties and solutions of artificial intelligence.
AI poet, assistant judge of artificial intelligence, winner of future Nobel Prize in Economics.
Language is an important symbol of human wisdom, and its position and role in human civilization are beyond doubt. Natural language processing, in layman's terms, is to "let computers learn human language." The newly established center in Tsinghua University aims to build a stronger knowledge base for machines by strengthening the basic research of artificial intelligence and its cross-cooperation with humanities and social sciences, so as to enhance the ability of artificial intelligence in language learning and processing.
Sun Maosong believes that there are two main bridges between artificial intelligence and social sciences and humanities. One is narrow data, such as data in financial markets, which can predict market behavior according to the data; The other is the text formed by language, such as news, messages, and various "short messages" that users reflect their views and emotions on social networks.
Based on the bridge between human language and writing, from the end of 20 15, he directed a research group with two undergraduates as the main body, and through the method of deep neural network, the machine can "read" and "understand" all ancient poems. On this basis, he also launched the AI—— that can write poems-"Nine Songs", which made artificial intelligence "cross-border" into the field of literary creation. Due to the public's concern and participation, Nine Songs has produced almost 4 million poems so far. "The total number of poems handed down from ancient times to the present is estimated to be less than 4 million. If you think so, the effect is not bad. " At present, the group is studying the emergence of modern poetry to meet the increasingly diverse "poetry requirements" of the public.
Social humanism is a very broad concept. In addition to literature, this year, the Institute of Artificial Intelligence is cooperating with the Law School of Tsinghua University to add more elements of computing and artificial intelligence to the original law. Law schools are increasingly aware that "computational law" is a development trend of future legal research. Sun Maosong also mentioned that the digitalization and openness of archives led by the Supreme People's Court in recent years is an important foundation for the development of computational law.
When our reporter (www.thepaper.cn) asked whether there would be "artificial intelligence judges" in the future, Sun Maosong said, "There should be" artificial intelligence assistant judges ". Although the ability of the machine is definitely superior to that of a good judge in essence, the machine does have its advantages. For example, it can read a large number of files quickly. People sometimes judge one-sidedly, their emotions will fluctuate, and the level of judgment will be uneven, but the machine will not. "
Sun Maosong pointed out that the artificial intelligence deep neural network method based on big data (also called deep learning method, see below) may bring about some changes in the research paradigm in the social and humanities fields. Although Tsinghua Institute of Artificial Intelligence has just started the research on foreign exchange rate forecast, he believes that the deep integration of artificial intelligence with finance and economy is expected to bring extremely powerful results. Once the deep learning method is creatively and systematically applied to this field, it may even "win a Nobel Prize in Economics in the world".
In view of the high complexity of the financial and economic system, Sun Maosong believes that when all kinds of relevant information are covered and there are superb means to deal with these information effectively, the predictability of this complex system will increase significantly. He gave an example, "The ant kingdom has its own set of operating rules, but this rule sometimes becomes very fragile, because it may suddenly come to a person and suddenly kick, bringing disaster to this kingdom. Just from the perspective of the world of ants, this kind of emergency is unpredictable. However, if we include human factors and carefully understand the nature of this person, such as whether he is a caring person or a Buddhist, the predictability of his behavior will be greatly improved and the unpredictability of the ant kingdom will be greatly reduced. "
In Sun Maosong's view, the research paradigm of classical finance or economics mainly depends on human rational thinking and solving equations (whether linear or nonlinear). However, the complex and huge financial and economic system is full of uncertainties, and many phenomena are difficult to be clearly described by equations. However, the deep learning method is highly nonlinear. Driven by big data, any complex equation can be simulated implicitly in theory, and the uncertainty of the system can be handled more effectively. This makes it possible to produce a brand-new research paradigm of finance or economics.
Turn "black box" into "grey box"
Yoshua Bengio, the "godfather of contemporary artificial intelligence" and winner of the 20 18 Turing Award, believes that the great changes in the field of artificial intelligence in recent years are attributed to the emergence of "deep learning". In a signed article entitled "Deep Learning: The Renaissance of Artificial Intelligence", he said, "In recent years, deep learning has become the most important force to promote the development of artificial intelligence, and major information technology companies have invested billions of dollars in this area."
In an exclusive interview with this newspaper (www.thepaper.cn), Bengio said that the existing natural language processing system has mastered a lot of vocabulary and language conversion skills, but it does not understand the true meaning of sentences. Machines "will make some very stupid mistakes, even without the understanding of two-year-olds." At the same time, he believes that it is of great significance to simulate the neural network of the human brain to realize machine learning technology similar to artificial intelligence.
Despite its great significance, the current deep neural network system has a major deficiency: given an input, the profound reason why the system gives the corresponding results is opaque to people. In this sense, it is basically a "black box". This will greatly reduce the robustness of the system and narrow the application scope of the system. When we delve into why the machine makes mistakes (for example, why the machine translation system translates a sentence like this instead of that), the machine can't give an explanation. "Anyway, I translated it like this, so you can use it."
In addition, deep learning is best at dealing with relevance: entering the word "chicken crow" will automatically associate with the sentence "the sun has risen". This reflects a correlation, but obviously the former is not the reason for the latter. Deep learning now basically has no causal reasoning ability. Sun Maosong said that deep learning only "feels" that two things are statistically related, and I don't know if there is a logical relationship between them.
Sun Maosong further explained that the research of natural language processing also faces the same problem. In fact, "interpretability" is the dilemma faced by the whole field of artificial intelligence and the frontier of international academic research. Researchers try to make this artificial intelligence "black box" at least become a "grey box". Sun Maosong believes that the construction of knowledge base may be a solution in natural language processing. Take the writing of ancient poems as an example: "Baqiao", followed by "Liu Zhe". Deep learning should be able to capture this knowledge association from the poetry database, but it is hard to say whether many other knowledge associations can be learned. However, if we consciously list all things related to Baqiao in advance, such as "Bashui, Post Station, ecstasy, heartbroken, Chang 'an, Eight Views in Guanzhong, Li Bai, Li Shangyin, Meng Haoran, etc." And establish a knowledge map, then the machine can consciously write poems according to this knowledge map, which is more targeted and more targeted.
According to Academician Zhang Yue, director of the Institute of Artificial Intelligence, artificial intelligence has experienced rationalism and empiricism in its development history. The third generation of artificial intelligence should be the perfect combination of the two. Solving the interpretability problem is one of its core research tasks.
Because knowledge is naturally interpretable, knowledge base is interpretable.
"Although both general courses and professional courses should be considered, the focus should be on general courses rather than professional courses."
On the occasion of the fourth industrial revolution marked by artificial intelligence and big data, many universities in China began to lay out artificial intelligence education. 2065438+In May 2007, the University of Chinese Academy of Sciences established the Institute of Artificial Intelligence Technology, becoming the first new college in the field of artificial intelligence technology in China to carry out teaching and research in an all-round way; In 20 18, Shanghai jiaotong university and Nanjing university successively established the artificial intelligence college and the artificial intelligence college, and Nanjing university also recruited the first batch of artificial intelligence undergraduates.
As a top institution of higher learning in China, Tsinghua University has always attached great importance to the cultivation of top talents. Established in 2005, the Computer Science Experimental Class (named "Yao Ban" because its founder Yao Qizhi is the only Chinese winner of the Turing Award) is the cradle of training top computer talents in China and even the world. First-class higher education has the responsibility to help computer science and artificial intelligence reach the next encouraging height, which should be able to bring subversive influence to this era.
So, is it necessary to set up an artificial intelligence undergraduate major? Should artificial intelligence become an independent department at the undergraduate level?
"The most important thing in Tsinghua is stability. If there is no special demand, it will not change. If it changes, it will become chaotic. The two groups may compete in disorder. " Overall stability and active exploration is Tsinghua University's policy for artificial intelligence undergraduate education.
This year, Tsinghua University has set up a new "intelligent class" outside the "Yao Ban", which is called "Tsinghua School Artificial Intelligence Class". On the premise that the structure of computer education is basically unchanged, this small-scale experimental class will conduct in-depth exploration on the cultivation of artificial intelligence talents, especially top talents. The work of artificial intelligence research institute is mainly located in the postgraduate stage, especially the cultivation of doctoral students.
Before this center, Tsinghua University successively established seven research institutions, including the Basic Theory Research Center of Artificial Intelligence Research Institute and the Intelligent Robot Research Center.
Sun Maosong said that the core strength of these eight centers mainly corresponds to the "initial heart" of artificial intelligence research, that is, the study of human perception and cognition. The most important channels of perception are vision and hearing, so the center of visual intelligence and auditory intelligence is established; Cognition reflects people's language, reasoning, logic and learning ability, so it has set up a basic theory center, a knowledge intelligence center and a natural language processing center. Further exploration is the related research of intelligent human-computer interaction, intelligent multi-modal information interaction and even robots. These centers are closely related and cooperate with each other.
Tsinghua Institute of Artificial Intelligence, which has just turned one year old, has basically achieved the goal of "integrating strength and making overall arrangements". Sun Maosong said that in the future, the two centers will use "social network big data" and "artificial intelligence chip" as keywords respectively.