The second phase of the lecture on artificial intelligence, hosted by the CAAI, opens in Beijing, January.18, 2018. Professor Jian Yu, director of CAAI and deputy director of the special committee on machine learning, delivered a wonderful report titled "Talking about artificial intelligence".

There are more than 300 primary and secondary school principals (educators) listened the lecture. The report in detail sorted out the definition of artificial intelligence and the "three schools" of artificial intelligence, and through the analysis of the basic problems now artificial intelligence is facing. Also professor Jian Yu answered questions for the audience.

Jian Yu: I am very glad to have the opportunity to speak a little about artificial intelligence on behalf of the CAAI. President Shiyi CHEN of the Southern University of Science and Technology has just given a very impressive report. As the next speaker, I am under great pressure. After all, this is the first time I've ever given a semi-academic presentation. I used to give purely academic reports. I will try my best to do it. If I can't do it well, please forgive me. First of all, let's talk about the report outline. Today's report is divided into five parts. Now let's talk about the first part.

Before we get to the first part, let's talk about why we're talking about artificial intelligence. Why would the Society for Cultivating Innovators (SFCI) let someone to talk about artificial intelligence? Actually, President Chen has already answered this question, because we have entered the era of artificial intelligence. I will deliver a more detailed discussion follows.

In terms of national policies, on July 20, 2017, the State Council issued the Development Plan for the New Generation of Artificial Intelligence, and some domestic and foreign countries like France, Germany and the United States also issued some policies.

From the perspective of the industry, nowadays, almost all IT enterprises declare themselves as artificial intelligence enterprises, like IBM, which used to do hardware, are now focus on artificial intelligence, and also many software enterprises such as Baidu, Google, Microsoft are all focus on artificial intelligence.

And in terms of practical products, that is too much. For example, AI chips such as Cambrian 1H8 and Baidu Apollo plan to open up autonomous driving platforms, these are just a few examples.

Even the ethics of artificial intelligence have been put on the agenda. For example, in January 2017, 23 Principles of Asilomar AI were proposed.

In order to illustrate the heat of artificial intelligence, here is a picture, this picture shows that artificial intelligence is not only hot in our ordinary people , but even monastic meetings have involved artificial intelligence.

When it comes to the origins of artificial intelligence, this report is not intended to go far back. The origins of modern AI are credited to the Dartmouth conference in 1956. The Dartmouth conference, which was attended by 10 people for two months, resulted in AI becoming an independent discipline. How to define artificial intelligence? There are two common definitions: "Artificial intelligence is a science that enables machines to do what people need to do with intelligence" , it’s put forward by Marwinsky. The second, more technical definition Nilsson gives is that "Artificial intelligence is the science of knowledge". What is the science of knowledge? It is to study the representation, acquisition and application of knowledge.

This two definitions are the most common definition of AI. Strictly speaking, there is no unified definition of artificial intelligence until now, but among the two common definitions, professionals prefer the second one. Why is that? The reason is also simple, because the first definition involves two undefined concepts, one is human and the other is intelligence. What is human? What is intelligence? It is still hard to say. The second definition deals only with an undefined concept called knowledge. Among the three concepts of human, intelligence and knowledge, knowledge should be studied thoroughly, and it is also the basis of intelligence.

Therefore, generally speaking, the research of AI is based on the representation, acquisition and application of knowledge, but all disciplines aim to discover knowledge. Therefore, compared with other disciplines, AI is universal, mobile and permeable. Generally speaking, AI plus a certain discipline can form a new discipline, such as bioinformatics, computational history, computational advertising, computational sociology and so on.

Therefore, to study artificial intelligence, we should study how to define knowledge. As for knowledge, all the Principals in this room, whether they are heads of universities or primary schools, will probably say that we know. Unfortunately, knowledge is also a cumbersome thing to define. Strictly speaking, the earliest definition of knowledge people used was Plato's definition: a statement that is proved, true and believed, or JTB condition for short. Unfortunately, this definition, which has been around for more than 2,500 years, was rejected in 1963 by a paradox called the gatier paradox, which I won't go into because of time. You can tell a story about a cow in a clearing, and through that story you can see that Plato's definition is wrong.

After knew that this definition was wrong, people later gave a lot of definitions of knowledge, which are still inconclusive. But one thing we do know is that the basic unit of knowledge is a concept, and knowledge itself is a concept. But if you really think about it, you will have a better understanding of artificial intelligence.

Knowledge is a concept in itself. Therefore, the problem of artificial intelligence now becomes how to study a concept, how to represent a concept, and how to apply a concept. Therefore, it is necessary to clarify the concept, which is also a crucial thing for artificial intelligence. Let's move on to the second part of the report, conceptual representation and reference.

When we talk about concepts, we're talking about classical concepts. If you think about the classic concept, many of the teachers in this room probably learned this during their undergraduate years. Let's briefly review that there are three representations of the classical representation of a concept in general, the first is called the symbolic representation: the name, what is the name of the concept, the second is the connotative representation, represented by the proposition, and the third is the denotation, represented by the classical set.

For example, prime Numbers. Its concept name is called Sushu in Chinese. Its connotation expression is a proposition, that is, natural number that can only be divisible by 1 and itself, and its denotation is a classical set, namely {1, 2, 3, 5, 7, 11,13,17... }. What do its various representations do? Or what does the concept do? It is easy to see that concepts have three functions or functions. To grasp a concept, all three functions must be clear.

Let's start with the first function, the referential function of the concept. The object function of the concept refers to the object pointing to the objective world, indicating the observability of the object in the objective world. The observability of the object refers to the perceptual and perceptual characteristics of the object to people or instruments, independent of people's subjective feelings.

Take an example from the true story of Ah Q: why did Zhao's dog look at me? In this sentence, "Zhao's dog" means a real dog in the real world.

The second is cardiac function. The mental function of the concept is to point to the objects in the mental world, and to represent the objects in the mental world. Here are some examples. Again, Lu Xun's article. Lu Xun wrote a famous article The Running Dogs of Bereaved Capitalists. Obviously, this is not a dog in the real world, but a dog in his mental world, namely the dog in his mind. In the objective world, Mr. Liang Shiqiu is obviously not a dog by any means.

The last function is the naming function. The naming function of concepts refers to the symbolic names that refer to the cognitive world or the symbolic world to represent objects, and these symbolic names constitute various languages.

Here is one of the most famous example of Chomsky, "Colorless green ideas sleep furiously", this sentence means a colorless green ideas in furiously to rest, it can’t tell us anything but completely conform to the grammar, it is said in the pure symbol world, recently little ice made some poetry of artificial intelligence, the sentence "It married many colors", in fact also meaning nothing but is grammatically correct. Of course, sometimes, there are other examples, such as in the sentence "鸳鸯两字怎生书", Yuanyang refers to the word "鸳鸯". Recently, a so-called divine comedy "Drunk Alone" also completely with the objective world, can be said to be a sign of the century's delirium.

In real life, if you know a concept, all three functions need to be right. You can't do that with just one. I'm going to give you a very simple example, which is a small story. One of my good friends is Shaoping MA. He is a professor of computer science in Tsinghua University. Once, he told me a very interesting story. One day, he went out for a meeting and ate alone at a table. Someone asked where is he working now. He answered Tsinghua University. After hear that the person was very happy, then asked, which department of Tsinghua University? Mr. Ma said he was from the computer department. The man said, "I know a teacher in the department of computer science in Tsinghua University. I don't know if you know him." Mr. Ma replied, "I have been in the department of computer science in Tsinghua University for 30 years. The teacher you said I should know, " the man looked up proudly and said," I know Shaoping Ma."

Did this teacher know Mr. Ma? Of course didn’t. He didn't know it was Shaoping MA who was talking to him. So, it takes three fingers to master a concept. If you can only refer to the name, can not refer to things, or can not be said to know the concept.

With these three concepts in place, it's time to delve into the genre of artificial intelligence. Now let's discuss the research route of how to make machines have artificial intelligence. Note that artificial intelligence is also a concept, and to make a concept a reality, it is natural to implement three functions of the concept. It is easy to see that the naming function of AI corresponds to symbolism, the heart function of AI corresponds to connectionism, and the object function of AI corresponds to behaviorism. Now, I’ll give a little more detail.

What is symbolism? The representatives of symbology are Simon (He Sima) and Newell, who put forward the hypothesis of physical symbol system. The basic meaning is that as long as the corresponding function is realized in symbolic calculation, the corresponding function will be realized in the real world. This is the so-called physical symbol hypothesis, which is the sufficient and necessary condition for intelligence. So as long as it's true on the machine, it's true in the real world. Say more popular a bit, the name is right, point to thing nature is right.

In philosophy, there's also a famous thought experiment, the Turing Test, on the assumption of a physical symbolic system. The Turing Test solves the problem of determining whether a machine is intelligent.

The thought Turing envisioned was this: Put a computer and a person in a room, both of whom were communicating with the outside world through a printer. People outside the room use a printer to ask questions to computers and people inside. The computer in the room answers separately from the person, and the computer tries to imitate the person. All answers are written in words by a printer. If the person outside cannot tell which is which, the computer is intelligent.

Obviously, the above tests are conducted at the symbol level, which is a symbol test method. This test has a number of advantages. Why? Because the definition of intelligence is actually very difficult to give, how to determine whether there is intelligence is also very difficult. But with the Turing Test, we can focus on the external functional performance of intelligence, making it seem achievable and judgmental in engineering.

The Turing Test limits the performance of intelligence entirely to naming. I have just given an example, only in the naming function to achieve, it seems that cannot be called to achieve. In fact, there are thought experiments designed to criticize the Turing test, based on the difference between naming and pointing.

This is a philosopher, Searle, who comes up with the famous Chinese room experiment, which is a refutation of the Turing test.

The experimental design is as follows: There is a person lives in a room, this person only knows English, but in this room there is a structured computer program, which can answer any Chinese questions, and the room has a window to pass out and pass in notes. Pass in the Chinese question through this window, the person in the room will output the corresponding Chinese answer according to this computer program. Apparently, the person outside the house thinks he is proficient in Chinese, but in fact, he knows nothing about it.

This is called the Chinese room experiment. The Chinese room experiment clearly points out that even if the symbolism succeeds, it is all about the calculation of symbols and does not necessarily connect with the real world. Nor is full naming necessarily intelligent. This is a formal philosophical criticism of semiotics.

Nevertheless, symbolism played an important role in the early studies of artificial intelligence. The major achievements of the early work of semiotics were machine proof and knowledge representation. In terms of machine certification, wang hao and wu wenjun made important contributions. The most important achievement of knowledge representation is the expert system and knowledge engineering, and the most famous scholar is Feigenbaum. There is clearly a problem with thinking that intelligence can be achieved along this path, and the subsequent failure of Japan's fifth generation of smartphones now seems entirely logical.

Specifically, what are the challenges to the realization of symbology? There are three main challenges. The first is the combinatorial explosion problem facing concepts. Because common sense is hard to exhaust, reasoning steps can be endless. The second is the combinatorial paradox of propositions. Both are reasonable propositions, and together they make sentences that cannot be judged true or false. The third is also the most difficult. Classical concepts are hard to get in real life, and knowledge is hard to extract. The above three problems directly led to the decline of symbolism.

Now let's look at the second doctrine of artificial intelligence. That's the mental representation of the concept, how to do it on a computer, that's what we call the mind, and connectionism focuses on that. School representatives include McCulloch, Pitts, Hopfield and so on. There was a very popular article on WeChat in 2016 about so-called semantic maps of the brain. The idea is that concepts can be represented in each brain region, and indeed there is a mental representation of concepts. How do you implement AI along this path? We won't talk about how. The philosopher putnam also criticized this line with a thought experiment known as the brain in a VAT experiment.

The experiment is described as follows: (one can assume that is you), underwent surgery by evil scientists, cut the brain down, in the cylinder of entities nutrient solution, the brain nerve endings connection on the computer, the computer in accordance with the procedures so as to deliver information to the brain for him, there are people, objects and the sky, feeling nervous and so on can be input, the brain can also be input, interception of memory (clipping off the memory of the brain surgery, and then input he may experience a variety of environment, daily life), input may experience a variety of living environment, and even can be input code, "feeling" to a funny and ridiculous that are reading this text.

Still, connectionism is by far the hottest AI implementation route. As President Chen has said, from neural networks to deep learning, AlphaGo beat lee sedol and then ke jie. It should be said that the research achievements of deep learning have made industrial-level progress. The trouble with this route is that the mechanism by which the human brain represents concepts is not well understood. The current neural network and deep learning are actually far from the real mechanism of human brain, and not the fundamental mechanism of human brain.

Now the referent, the behaviorism of AI. In a line, it is assumed that intelligence depends on perception and action, and requires no knowledge, no representation, no reasoning. It means something, as long as it can be expressed. The representative of this school is Brooks intelligent robot.

Philosophers also designed a thought experiment to criticize this doctrine, the so-called perfect pretenders and spartans. Both the perfect pretense and the Spartan performance have nothing to do with the heart. How is such intelligence tested?

The greatest difficulty of the behaviorist line can be illustrated by moravec's paradox. For computers, the hard problems are simple, the simple problems are hard, and the most difficult to replicate are the unconscious skills in human skills. There are big challenges in simulating human action skills. Some people say we've seen boston-powered robotic mannequins on the Internet. But it has a big disadvantage is that the energy consumption is too high, the noise is too large. Originally, the big dog robot was ordered by the us military, but it could not be used by the military, because the starting sound could be heard ten miles away. How could the military use it? That makes him a sitting duck.

Now let's talk about the development trend of artificial intelligence. Let's talk about the development trend of artificial intelligence. Semiotics says, it is enough to realize the naming function. Connectionism says that it is enough to achieve the function of the heart. Behaviorism says it's enough to function as a referent. However, the premise of the three is that the name, the object, the heart is equivalent.

But the question is is this premise true? Early AI assumptions were equivalent. Since the early AI used classical concepts, there are five hypotheses for classical concepts: the first hypothesis is that the internal and external names of the concepts are consistent. Secondly, there is a classical set representation for the denotation of concepts. Thirdly, the connotative representation of the concept is existential proposition representation. Fourth, there is a unique representation of the concept, that is, the representation of the same concept is independent of the individual, and everyone says this is the same. Fifth, connotation and denotation are equivalent. Obviously, under the above five assumptions, the functions of the heart, object and name of the classical concept are completely equivalent.

However, the concepts used in daily life cannot guarantee the functional equivalence of the heart, object and name, and the five assumptions of the classical concepts are generally no longer true. An example is given below to illustrate that in daily life, the naming and referential functions of concepts are not equivalent.

WeChat once passed a famous joke: a person said he had a billion, who has a project notice, together with investment. Otherwise, later than that, I'm done. People who listen to it think it means something, there is really a $100 million fund. In fact, the man had just three words on his hand: "one hundred million". In this case, the name simply refers to the symbol "a billion". This passage clearly takes advantage of the fact that the nomenclature of a concept is not necessarily equivalent to its referent.

Once there was a famous painting in the west, in which there was a pipe, but the inscription said that it was not a pipe. Here, it is clear that the symbol is different from the real object, that is, the name is not equivalent to the object. In real life, we can also find examples of naming and fingering are not equivalent.

To sum up, the naming, object and heart functions of concepts are not equivalent in life. Therefore, the realization of a function of concepts alone cannot guarantee the existence of intelligence. As a result, current ai advances no longer follow a single school of thought. Following a single school is not enough to achieve artificial intelligence. The current development of AI is comprehensive. The knowledge map developed from the expert system has not completely followed the route of symbology, and the driverless technology is a comprehensive technology that breaks through the limitations of the three major AI schools. These are time-limited, so we won't get into them here. As the former President Chen said, the new generation of artificial intelligence will bring about the fourth technological revolution of the society.

In the last part of the report, to supplement a little artificial intelligence common sense. It should be said that the current artificial intelligence is still very big defects. Because the basic concept is still classical. However, the basic assumption of the classical concept is still that pointing to the heart and naming are equivalent to pointing to things, which is seriously inconsistent with the common sense of human life. In our real life, it is not equivalent for the concept of naming to refer to the object and the heart. How to break through the above frame is still a very big problem. Under the above framework, the machine sometimes appears to be extremely retarded, lacking common sense, understanding ability, and sometimes seriously lacking understanding ability, which is especially reflected in the man-machine dialogue system, which shows that its dialogue viscosity needs to be improved.

In addition, I will tell you some common ai mistakes. Often read some of the literature, reports, talked about the so-called general artificial intelligence, strong artificial intelligence, human level AI, super artificial intelligence. It should be said that these concepts are wrong.

Why are these concepts wrong? Quite simply, assume that one of the above artificial intelligence is implemented. Once implemented, these concepts immediately face logical paradoxes. If it's not clear, think about who implemented them. If people have implemented them, aren't they already smarter than or at least as smart as people? Recently, zhou zhihua, a teacher from nanjing university, wrote an article about strong artificial intelligence. Interested teachers, you can have a look.

In essence, intelligence itself has layers. In Go, Alpha Go is already unique, but its intelligence level is not higher than that of the maker of Alpha Go. If so, the so-called universal artificial intelligence, strong artificial intelligence, human level AI, super artificial intelligence is wrong. There is an extreme AI theorem in the book, which is called tesler's theorem: "artificial intelligence is something that has not been done yet". Does the artificial intelligence, is not the intelligence, how surpasses the human or with the human same?

Why must strong AI be criticized? Because they have not retired from the historical stage as historical terms, they are extremely deceptive. They not only mislead the general public's understanding of artificial intelligence, but also have the market in some researchers and some government research reports. The following are some recent examples, such as the "2017 trend report on Artificial Intelligence in China", "Perspectives on Research in Artificial Intelligence and Artificial General Intelligence Relevant to DoD", etc. If you are interested, you can search some online. If we knew that strong artificial intelligence could not be established, the so-called female robot Sophia cheating incident would not happen.

Thank you. That's all for my presentation.