Hello everyone, Zhongzhi SHI and Feiyue WANG are both senior teachers in this field, because I remember when I was studying as an undergraduate in my artificial intelligence class, I used Mr. Shi's books as reference books. Mr. Wang is also a very outstanding expert.
Today I feel like being a student, learning something new. I learned that to be a good scientist, what we need is not just to study technology, but to study philosophy, which means we need to study the nature and laws of matter, and to study some of the most basic things, starting from the basic and then moving forward.That's why Marvin Minsky is one of the people who has had a big impact on us. Because when I was pursuing my PHD, artificial intelligence was at a low point. At that time, people who used to study AI didn't do it anymore, instead, they came to IA, in which A was referred to as Agent which SHI mentioned before.
I was at Tsinghua University to apply for a PHD in an AI lab. If I had applied to join the network group, it would have been a disaster. There’s a change, and in a certain degree, Marvin Minsky has established himself as one of the most influential people in this field because his book named Perceptrons could have silenced connectionism for decades. In fact, it's only now that some of the work of deep learning is really starting to come into play. I think a lot of studies would have results a decade or so earlier if we had just moved on at that time.
From another perspective, when God closes a door, he will also open a window for you. Many developments in other aspects of technology have emerged. In fact, it should be said that the development of science and technology is a spiral process, it has no end. Just like a hundred years ago, it is difficult for us to imagine what will we have today. The Industrial Revolution brought us telephones and televisions, and we thought no new technology was needed in the future. But we now find that these technologies are still developing with each passing day, so the technology is advancing in a spiral way.
Then we look back and find that the whole history of AI is also such a process, it has ups and downs. From the highs and the lows, we see these very fresh ideas that have actually driven the discipline forward. You know, at the Dartmouth conference, Marvin Minsky was 29 years old, just a few years out of his PhD, and it was a very young group of people doing this.
In fact, when people are working on computers, they think that computers can replace or solve some of the problems of the human brain. Machine Translation, for example, is a problem that people find difficult because language is one of the great differences between human beings and higher animals. In 1954, a computer and humans collaborated on an experiment. At that time, a demonstration was made. More than 60 sentences in Russian were demonstrated and automatically translated into English. At that time, it was said that the robot translated automatically. You can also read the translation results. I don't understand Russian, but I do understand the translated English version clearly.
You would think that a machine could solve a problem like machine translation, maybe in a few years. Because you see how well a machine can translate, and these 60 sentences are from different fields.
But when actually doing machine translation, we encountered many, many problems. The IBM research department has been doing this for many years and is still doing it. We've seen constant improvement, and just like with AI, it's a gradual improvement. Many approaches based on machine learning have been proposed, and tasks related to machine translation emerged. To make better use of big data, a lot of Internet companies have been using data to solve translation problems.
From this point of view, we have the dream that in the business field, will the communication between people be well solved by artificial intelligence technology?
Besides, recently, AlphaGo has also made us very excited. Chess is a great manifestation of intelligence, because playing chess shows a person's IQ, thinking mode and computing ability.
In fact, if you look back in history, 40 or 50 years ago, or even a little bit earlier, a lot of algorithms were invented. IBM tested it in 1994 with checkers. If you look at DEEP MIND, which is the company that AlphaGo belongs to, machines could randomly read the pixel characteristics of the game. Although the machine does not know what the rules of the game are, and only plays some simple games and give feedbacks, after a few days, it could play these games very well. This machine learning capability is actually based on testing done by IBM in 1994.
A lot of times, the early tech guys had good ideas that solved certain problems at the time. People may slowly forget about it, but in retrospect, when new devices are developed, including some big data, it may be hard to remember when it reappears. Simon's checkers program was also a hit. It was televised at that time. The game won a state championship, and IBM's stock rose more than 10 percent the next day. But after decades, it was forgotten.
This is how the technology continues to develop in a spiral way, and its overall trend is upward.
What new changes are there in the industry today? When we look at today's technology, when the word artificial intelligence was put forward 60 years ago, or when Marvin Minsky wrote the book perception, what changes have taken place in these decades? In fact, if you look at a lot of technical changes, these algorithms, concepts and directions have been well defined at that time. It's just that many aspects are not so detailed.
But now there is a new change. We have really ushered in an era of big data. Data has become a kind of resource that can be used, and it is growing continuously. Its driving force comes from several aspects. On the one hand, it comes from the development of mobile technology. Equipment generates data, including data generated by camera head. A mobile phone can now generate a lot of data, which can be broadcast by mobile phone. Now a lot of such data are generated every day. This data is not available in many scenes before. I remember watching Mr. Li Feifei's net, which he did. I thought that I could build a data set that is big enough to be labeled. There are also many visual identifications. Children can get a lot of data at 25 frames per second, and the data is marked. His mother told him that this is an apple So he has enough data. From this point of view, if I have enough data, I may be a good foundation for pattern recognition and image recognition in the future, so he has built an environment for many resources to mark. When the mark is large enough, there is a quantitative change to a qualitative change.
In addition, there is another person's data. My mobile phone is broken. One day, I tried the era of no mobile phone, and my work efficiency was very high. Without WeChat, I could work wholeheartedly. However, I didn't eat a meal at noon that day, because I usually ordered meals, and there was no card in the canteen, so I bought some biscuits. I found that the biggest problem is that I can't take a taxi. IBM Research Institute is also in a relatively remote and quiet corner. There are not many buses around. I used to use Didi or Uber before taking a taxi, but I really can't do without a mobile phone. Now this society has become a mobile society. We used to hope to build an agent. What is this agent? He can understand me. He can communicate with the physical world and the computer world on behalf of me.
But in fact, this agent has already appeared, that is, your mobile phone. When there is no mobile phone, I will feel very painful on that day.
This is because when there is a lot of social media like this mobile phone, it also brings a lot of data, which is actually terrible data. In the future, let's think about it. How many things could a person write before? It's hard for you to imagine. When you think about it, you put it down. How many letters can you write in a day? It's hard to write. If a post-90s or post-00s child puts his life online every day, sends a post every day and saves them together. He can write several books every year. The key data are all electronic. From this perspective, he stores all the information. And these data help us to solve the computer algorithm in the past encountered a challenge? I don't have enough training data. For example, there has been a great breakthrough in deep learning recently. It has been proved for a long time that there are layers of neural networks. How many layers of neural networks are good? There was no result. But the main result is that in such a large parameter space, each neuron has several parameters, and the connection between them is all some parameters. These parameters need enough data to help him and avoid him. However, due to the emergence of big data and the improvement of computing power, we can easily get a (English) software and put it in your computer. If there is a GPU, it can be accelerated.
These technologies are very difficult for our traditional technology, such as image. When I read PhD, my instructor said that image is a good topic. You can do it for 50 years. 50 years is a long time. You can do it for another 50 years in a while, because the image is very difficult. Imagine that the computer has captured an image. If you want to understand the image, the first step is to do image segmentation. Many subjective assumptions, I do feature analysis, in this hypothesis, in fact, there is a lot of noise, we have done a lot of work in the past, found that in the paper, the effect of some scenes is very good, in the real scene, it is difficult to have the same good results.
I remember that I did a paper on image retrieval at that time, and my value was calculated to be more than 20%. At that time, my instructor said, "Suzhong, you are very courageous.". Because you see a lot of papers published, a lot of the basic line is 80%. Deep learning is the process of simulating the brain from the structure. A brain does the process of vision. From this perspective, if we give it enough data and have enough computing power, it can do better.
In fact, some time ago, I met colleagues from Microsoft. They have already done more than 3.0% image recognition. What is this concept? How good can people do? The error rate of human is about 10%. On that basis, its image recognition has been better than that of human beings. It is not an ordinary person. His image is multi class image, not one kind of image, there are many types, as we generally recognize some ordinary ones.
In image, voice, we now such technology, because of the occurrence of big data, it has achieved a greater breakthrough. What is the result of this breakthrough? We just talked about the emergence of big data, especially the emergence of unstructured data. Computer science and technology understand that it is a big difference. Because the computer is a calculation, it can calculate faster than people. If what is done in the business field? Process automation. It's a good calculation. For example, the money you have in the bank is calculated for you. And if you look at our interest rate now, we can calculate the daily interest rate and even the small interest rate, because the cost of calculation is getting lower and lower, and the calculation speed is faster and faster. I remember my mother used abacus to save money at that time.