Zhou Zhihua, Professor, Nanjing University, Deputy Director of the Department of Computer Science and Technology, Deputy Director of the State Key Laboratory of Software New Technology, Director of the Institute of Machine Learning and Data Mining (LAMDA); ACM Distinguished Scientist, IEEE Fellow, AAAI Fellow, IAPR Fellow , Fellow of the Chinese Computer Society; Cheung Kong Scholar Distinguished Professor, winner of the National Outstanding Youth Fund. At the 2016 CCF-GAIR Global Artificial Intelligence and Robotics Summit, Professor Zhou Zhihua accepted an interview with Lei Fengnet (searching for the “Lei Feng Net†public number) and shared his views on machine learning and the future. For the general public, the concept of deep learning may be very popular. In fact, only 9% of the top academic conference NIPS 2015 (Editor's Note: Neural Information Processing Systems, one of the best neuroscience computing conferences) in machine learning last year came from deep learning. The total number of relevant papers is 11%. About 10% . Although only 10% in the research community, it may be 90% of the public view. I think that in the future, there will inevitably be other technologies that may meet or exceed the current status of deep learning technologies. Neural network learning is a branch of machine learning. It is particularly popular because of its great success in images, video, and voice . And these aspects happen to be more familiar to the general public, and this has caused everyone to have this kind of view—feeling that “neural networks are particularly hot.†​​In fact, there are other machine learning methods that have also achieved great success in their respective fields, but the general public Little is known about the relevant aspects, so you may know more about neural networks. In fact, many of the machine learning are the same, so-called deep learning - to see it as a language, it is not so completely different from the previous method. Think of it as a description of the way before a lot of content we describe it this way. In fact, today's field of deep learning has been integrated into many mechanisms of machine learning in the past. In fact, they themselves have been figured out, including some common theoretical problems, and they are all the same. In fact, the entire field of artificial intelligence in China should be said to be one of the most consistent with the international level in the computer science community (it can be reflected from many indicators). If China's relevant research level is comparable to that of foreign countries, our most powerful field is actually no different from foreign countries. However, our study may not be thick enough. For example, we may have reached a relatively high position in one aspect, but there may be a lack of other aspects. After all, it is still several decades later. The next step in machine learning is actually going in all directions. One big trend is that because of the increasing number of different types of data, there will be more and more data that needs to be analyzed. Every new task requires a new technology, which can be said to be a state of radiation . Many tasks may require new machine learning techniques in the future, but a big trend is to increase the robustness of machine learning . This is a very big demand, because in the current situation many machines in the research can reach human standards, even better than the human condition. But if you encounter some rare situations, it will be very wrong. In some applications, this is a situation that people are very reluctant to see. For example, driving without a driver is better than normal drivers. However, if the performance is not good, the consequences will be incalculable. Although it is technically possible to send all kinds of radiation out of all directions, there are still many other areas that need to be done. This view is not quite right – because after increasing the number of layers, the model is more complex and can eat more data. But after eating it, will the performance of the model be better? This is not necessarily down. If you only need a hundred layers, you will be one hundred and twenty layers. In fact, performance will worsen. The complexity of the sample and the complexity of the model must be just right. In terms of parameters and learning theory, it can be done, but there are many "trick" mechanisms in the neural network. Many people tried to do it and tried out many different things. However, the threshold for theoretical analysis is very high. To find common ground, we can do theoretical analysis. Now everyone is blindly trying, the results are also good, so the corresponding theoretical analysis can not keep up. Computer science is an application-driven research. It can be said that China currently has several leading Internet companies, communications companies, and some multinational companies. We all have cooperation. Usually, when they encounter some data analysis problems, the existing methods cannot solve them, and we provide a set of solutions. But from the perspective of the public, he is still the original financial companies, firewall companies, driving companies and so on. Nowadays, the connection between industry and academia is much stronger than before. One important reason is that domestic IT companies have developed. It can be said that some of the results we learned from machine learning ten years ago have been very useful, but the industry did not have this demand at that time. In fact, in terms of deep learning, the industry can be seen as eating and drinking tomorrow. It is after considering the issue a week later. At this time you only use these technologies. In fact, it can also be said that when the economy develops to a certain extent, and the enterprises develop accordingly to a certain extent, these technologies may be more useful. Once it is found to be useful, it will lead some firms to follow-up, which is actually a driving process. In fact, neither praise nor criticize deep learning. This is a very natural process of technological development. After five years or ten years of machine learning, there will be a new kind of technology that has become very popular at the time. For example, statistical learning in the 1990s, probability in 2000, and deep learning in 2010. I think that the biggest problem in the field of machine learning is that basic theoretical knowledge has not kept pace with, and more people are trying to do it, and lack relatively strict theoretical knowledge. Silicone USB Cable,Micro USB Type,Micro USB Cable,Lightning Data Cable Dong guan Sum Wai Electronic Co,. Ltd. , https://www.sw-cables.com
Deep learning technology in machine learning
| The Craze of Neural Networks
| Some progress in current deep learning
| Where is the gap between China's machine learning research and the United States?
| What's the next step in machine learning?
In practice, many deep-learning research scholars have jokingly called "not enough performance, add money to join in." How do you think of this?
| Is there a way to quantify complexity in deep learning
| Application of research results
|The current connection between academia and industry in machine learning
| Praise or criticize machine learning in one sentence