Tencent in terms of AI is indeed very low-key, a lot of people ask me Tencent has not done AI? How come it never comes out?
Decryption Tencent AI Department
In fact, from the beginning of 2016 in April, Tencent set up its own AI department. At present, the Department has more than and 30 scientists, more than 90% of people are doctoral degree, the vast majority of people are returned from overseas universities, including Harvard, Connell, MIT, Columbia University and other colleges and universities.
Tencent's AI may not be as famous as other giants, such as Google, which owns AlphaGo, Baidu with a car and a degree of mystery, and Facebook with some of the best image recognition programs.
Tencent AI is mainly concentrated in four vertical areas: computer vision, OCR, voice recognition, NLP. Each area will be expanded to more in-depth research. In the field of computer vision, in addition to the traditional image processing, but also related to AR and spatial positioning technology.
Speech recognition, in addition to the traditional speech recognition, speech synthesis, will also be involved in the relevant technology of automatic translation. In addition, in addition to the traditional cognitive behavior of Natural Language Processing's research, Tencent also these technologies for the development of floor chat robot, etc..
Tencent AI research direction of these four, and Tencent's existing business closely.
Tencent as a social long company, according to social networking services to create AI capabilities and products, including chat robots, intelligent assistants, will be based on social research.
At the same time, the game is a very important business Tencent. The future will introduce more Tencent in the game AI, imagine a scenario, the future is not LOL AI can participate in this competition in the world, with other people. Tencent also has a very popular hand tour called King glory, if the use of AI to enhance the ability of this inside, it will improve playability and fun.
In addition, Tencent will provide us with a lot of tools such as AI, including face recognition, speech recognition, Natural Language Processing, as well as learning platform.
Three reasons for machine learning to make a breakthrough
AI in the past 60 years has been in the last year of ups and downs, then suddenly broke out, until now.
In 90s, Li Shishi defeated Kasparov in the deep blue and won the championship in the dangerous edge challenge, then beat AlphaGo last year. There are a lot of technical aspects of the evolution of the entire history of AI, the first is the breakthrough in deep learning in 2006. Early man wanted to learn how to fly. The original method was to feather feathers, like birds. But in the end the real principle is to fly through the air to solve the principle of birds, which is a deep learning. The reason why we can now in a lot of this industry, in a lot of applications to achieve a breakthrough on the above, in essence, is to grasp its internal methods, rather than the surface of the method. Artificial intelligence is also the case, the study of learning algorithm is very important.
The second is the promotion of the model. Machine learning in 80s and 90s is also very fire, there was a SVM, it is already a very powerful machine learning algorithm. When it reaches hundreds of millions, billions of scale, in fact its computing power will drop rapidly, a composite function is very complex to describe this way, it is through the BP multi connection, reach an exponential relationship between the layer one billion times, the description may only need to connect one thousand nodes of the three layer, can build one billion character. So from its own point of view, the model is also a breakthrough to enhance the depth of learning.
Third is in the mathematical level, that is, BP above the problem of solving the problem of back propagation. In the mathematical theory firstly, back propagation is a very complicated problem, when a lot of things in the transfer layer network in neural network, when it back we need to reverse the convergence, are going to approach the best quality. However, when the number of layers is often too much, there will be a gradient disappear or gradient expansion problem. Through a number of methods to solve the mathematical theory of the solution, a good solution to this problem, so in the mathematical theory, the establishment of a better foundation.
It is precisely because of these three advantages, so that the machine learning so fire. But this wave will continue far away, in 1933 to 2000, the whole process of learning traditional shallow machine learning, it has a good descent, but the 2000-2010 ten years, its progress is very fast.
In the methods and models may not have been carried out, suddenly in 2012 or so, Microsoft Institute, they began to use the machine learning in the industry to identify the voice inside, and made a great breakthrough. The whole process of machine learning is indeed in the past five years, the development is very fast. Talked about a lot of machine learning this, just said the development of rapid development, it is also a good method, the model is also very good, mathematical algorithm is also a breakthrough, but what is the status quo?
Limitations of artificial intelligence
Although I am very much looking forward to AI, but also that there are many limitations of AI.
The first is the ability to learn. AI compared with people, it has a big gap. In fact, now all machine learning methods, we have found that no matter how this new proposed learning process it is to start from the beginning, the data must be re conduct a training process.
But there is a big gap compared with the ability to learn, people have a lot of intelligence is innate, born just like a child, don't need long time to know the world is three-dimensional, when you put a thing on a plane, put a bottle on the TV and he knows, there is one thing behind the TV, there are some features that are innate, this is related with biological evolution is.
Therefore, this kind of human, primate animals compared to animal cells, is certainly a natural ability. However, the depth of the learning method is very regrettable, scientists no matter how good the model is put forward, it is necessary to start learning, this is the first with the ability to learn, the machine is very large defects.
Second is no matter how good the learning model, it is essentially through the calculation, the use of computing power to solve big data, with greater computing power to do a better convergence. The past is the development of hardware, is obedient to Moore's law, development is very fast, but the following parameters in more in the future, people still have this ability to calculate the effect of this play on a big question mark.
For example, the network model proposed in 2006, to the back of the model proposed by University of Cambridge, and then to Google, and then to the neural network in 2015. Each new model is proposed to increase the number of layers of the model, the neural unit is more complex, the training results are longer, the results obtained are also optimal. But this method is not as sustainable as the original method, it is a big question mark. Another is to solve the problem of the image, the following to solve the problem of perception, if you want to solve the problem of cognition, the gap is even greater.
Human language is a sequence of problems, the language sequence problem if you want to calculate, this calculation can not be solved anyway. It is easy for a person to go back to a keyword in a fragment of a long sentence, but it doesn't necessarily work in a machine. Although the length of the memory element model formed in the first model, now Tencent, with attention to the model, but in short terms, the evolution of this model are also compared with people is very complex, is far better than people.
For example, that day I saw a dialogue, three people in the dialogue, the two people in the chat, there is a large section of the middle where to eat, suddenly someone asked the sun? People know that this is the description of the Suns, because in front of long before someone in the chat, the Lakers and Clippers topic, suddenly everyone knows to the sun, but the machine is basically no way to identify the. How much can wear to wear in winter and summer, how much can wear to wear, the two are basically the same, but go ahead to the description, sticking out, behind the winter show.
The second example I'm talking about is speech recognition, I saw a joke, speech recognition is difficult to engage, Hello, convenient interview? When I repeat this sentence, I do not know about it or is convenient for the interview, instant noodles, this is indeed a very difficult problem, but there are a lot of people's consciousness, because when asked, then slowly bring out this thing. So AI said the current situation, in the image, including face to do how badly, in fact it in a lot of constraints, it may not be able to face by face recognition, when wearing a hat is also difficult. So it is under a lot of constraints, the face recognition rate may reach only 9%. These problems are also reflected in the speech recognition, speech recognition is also relatively small noise, no wind noise, car noise environment, the machine recognition rate will have a better effect. But compared with the basic ability of the machine is still a big gap, not to mention it in terms of cognition, in terms of NLP.
How to improve the ability of machine learning?
AI with people, including some of the gap and machine learning ability, how to make up the gap?
The first one I think compared to people to create, we are now all based on big data, these data come from, it is very important. This data is now to get the traditional, but more data is itself can be created, of course, this method by just introduced, including AlphaGo has been verified this problem, through reinforcement learning to humans has never been to the game, this is a kind of creative ability of data, generate more data through the creation of data capacity, not necessarily create these, to make these things more exhausted out. In the future, if you develop in this area, it may be to enhance learning, to make more development and breakthroughs.
The second is what is called by analogy, analogy, AlphaGo can go down the world champion, but his approach to playing chess now has to die. Because his method is not created for playing chess. When we verify large data and a very good effect, when we have a small amount of data, how to migrate the original model, which is a very important research direction.
The third direction is compared with people is summarized. People are very can make some summary, including Newton's first law, including Newton's gravity, are summed up, including many of our axioms, but there is no way of machine learning are summarized, we have good results, it does not refine the relationship of justice and law. So in the future we want to summarize, especially unsupervised learning above, the classification problem is a goal to learn, but the clustering problem is not the goal, how do we get together.
So in the above three abilities, I believe this is our future in the AI above to ascend. The second is in the direction of the whole of the above, just speak a lot of said machine learning through the data from the traditional method, it is shallow learning inside, we all use statistical probability theory, statistical probability theory complete to support it, we ask the question when it comes to the limit of function, there are many mathematical formulas to complete that can solve its problems, but just learning on our machine, although the use of random to find local optimum in front of us, but it is the field of mathematics, it is only a framework, we are still a heuristic constraint in many of the above, we include how much is the initialization parameters, including our learning rate this is instructive.
In the future if the machine learning to continue to go down, we must have a strong support in the theory of mathematics, especially the traditional machine learning, mathematics, to migrate to our machine learning.
We know that neural network is put forward, many from the original brain science and biological systems, in itself this above, I believe that the introduction of more to the future development of AI, we are not the only mathematics and computer science, but also the introduction of brain science nerve directs, because the neural structure of the brain is, I just mentioned the cicada yarn neural network, the concept of brain nerve has been connected, cross layer connection, to achieve a very good effect.
Of course, more things I hope in the future, in the cross disciplines above, including the integration of biological, brain nerve and even philosophy, so that the whole AI may have more complete development. Another place is that in today's world, AI for all companies, for all people should be equal, so we must open. Good is the world all the big companies in the field of AI which are doing, will also include Tencent, including Google, we see a lot of FaceBook it is open network model very much, including we know now is the fire Open AI etc. these large this machine learning platform pioneer. Tencent in the future will go to a lot of open, so that more people to participate in the test. So the whole process in the future development of AI, I believe that the first is in the ability to go above, to go with the person to match, to enhance. Second in the whole study of the complete, complete mathematics, discipline on the rich. The third is that all large companies should be more open to the face of AI, which is the future of AI, AI's future.
China's AI strength no less than the United States
Here I would like to emphasize once again, AI is very important for Tencent, China's Internet is very important.
In the Internet era, we go to the United States than the strongest companies, there will be some gaps. But in the AI era, including Tencent, including the major Chinese companies, is completely comparable with the world's top companies. Why?
First, we have a large amount of users, enough data.
The second is the application of the scene, as the Tencent is concerned, we have a lot of AI this remote technology with the opportunity to fall, we can WeChat, games, news, QQ to go inside the scene AI. Even a little bit of speech recognition, image, chat robots, etc., which is where we can go to the ground.
The third place is talent. At present, there are a lot of Chinese in the field of machine learning, I participated in the 2016 ICML, there are three thousand people, I dare say about 30-40% of the people are chinese. 40% of the articles are written by chinese. The personnel structure is the basis of the China talent is very good, that is based on the advantages of our data, we drop our scene advantages, the structure of talent advantage, I think in the Tencent or other Internet Chinese have a brilliant future in the future of AI. Finally talk about Tencent's AI mission. Our AI mission to make AI ubiquitous.