Today, the fourth paradigm has released a “prophet†for artificial intelligence development platform, which is the first development platform for developers in the AI ​​field. The fourth paradigm sets parameters automation algorithms for the prophets and builds a machine learning infrastructure that is hundreds of times faster than Spark. It reduces the manual labor involved in feature engineering and model training, and can provide automatic or semi-automatic feature engineering. The model selection parametric tools reduce the dependence on data scientists.
The prophets can provide better application-oriented solutions to practical problems, and practitioners can use this platform to become data scientists within 1-2 months. Founder Dai Wenyuan joked: "Our vision is to make our own scientists unemployed."
The fourth paradigm calls for "AI for everyone" and wants to break the restrictions of the AI ​​circle. Dai Wenyuan said: "We are not trying to set up a small circle. We, these people, play AI in this small circle, but we want the AI's threshold to be lowered so that everyone can participate." And the fourth paradigm chief scientist Yang Qiang It also regards "assuming artificial intelligence is monopolized by certain groups and guiding artificial intelligence to the public" as its mission as a scientist.
At today's press conference, the fourth paradigm invited Li Kaifu, founder of Innovation Workshop, and Zhou Jiangong, chief executive of First Financial, and Shen Nanpeng, the founding partner of Sequoia Capital, to delay the flight due to bad weather, but also halfway through the press conference. When they arrived at the venue, each guest was prepared and gave a keynote speech. Coupled with Dai Wenyuan and Yang Qiang, the conference’s specifications will support a quarter of the artificial intelligence summit.
After the conference was over, Lei Feng and Professor Yang Qiang, the fourth paradigm chief scientist, discussed more about the “prophet†and migration learning. Prof. Yang Qiang is the Director of the Department of Computer and Engineering at the Hong Kong University of Science and Technology, the first American FAIlow of the AI ​​Association (AAAI), the only AAAI Chinese Councilor, and the chairman of the international top academic conferences such as KDD and IJCAI. He once founded Huawei’s Ark Laboratory in Noah’s Ark. Chief scientist.
Lei Fengwang: Prof. Yang Yang, you said that the data is that capital is oil. People with data can provide more and more services. Does this mean that companies in the AI ​​field are still left to large companies?
Yang Qiang: In fact, any successful artificial intelligence application is inseparable from self-learning, there are enough data, there are enough needs, huge computing resources, and there are top data scientists to build the system, these are necessary conditions, It is essential. If we go like this without any new measures, technologies and platforms, we will see that the rich are getting richer and richer. What does rich mean? There are more and more scientists; what do current university professors continue to resign? Go to Google; data continues to accumulate, where to accumulate? Google, Baidu; formed such a "rich richer" situation.
This is the current stage, because only large companies have the ability to collect this data, and this phenomenon will become more and more serious. He has the ability to collect data. His data is more and more. He can generate new services through data. In turn, these data can attract more data, so the snowball is getting bigger and bigger. This is not a very healthy thing. . We also want to break this point, so that companies that are not big companies and have not had such great financial resources can enjoy the benefits of artificial intelligence. We call AI for someone. For some people, AI can be profitable. This is an improvement over the previous AI for no one. Our philosophy is AI for everyone.
Lei Feng Net: Is there a hope for unicorns in this area?
Yang Qiang: I think the unicorn basically has appeared. The following will depend on whether the unicorn will continue to appear. For example, Google is a unicorn, because the search itself is a combination of big data and artificial intelligence, including Google now The products are driven by big data. In addition, we are seeing drabs as a data-driven company because only data can better coordinate and schedule them. Like Alibaba is also an artificial intelligence unicorn company, because it is the use of a large number of e-commerce data to do a better recommendation platform and transport platform. In this case, its service will become better and more efficient.
Lei Feng Network (search "Lei Feng Net" public concern) : just on the stage Dai always mentioned that there are some limitations in deep learning, the limitations of performance in what place?
Yang Qiang: It is true that deep learning has some limitations. Its limitations come from several aspects, because a model is a reflection of reality after all. It is equal to the mirror image of reality. The more capable it is to describe reality, the more accurate it is. However, we see that there is a limit to deep learning because machine learning uses variables to describe the world. The number of variables in deep learning can be limited and the depth is limited. In addition, the demand for data increases with the model. Larger and larger, there are not many situations where we have such large and high-quality data in reality. In fact, on the one hand, it is the amount of data. On the one hand, there are variables in the data. The complexity of deep learning to describe the data is not yet complex enough. Therefore, there are still limitations in this aspect.
Lei Feng network: But deep learning it is not the best method of machine learning?
Yang Qiang: It should be said that at present, it is the best for certain problems, such as face recognition and speech recognition, but it is not the best for other problems, such as delaying feedback, such as the action of robots. AlphaGo is not a game that is played in the depth semester. It also has part of reinforcement learning. Feedback is that you don’t know your winning or losing until the last step. There are many other learning tasks that are not necessarily deep learning to complete.
Lei Fengwang: With regard to the current level of data analysis, can we accurately predict the outcome of the World Cup or the European Cup? I saw a team predicting the results of the European Cup before, and I feel that it seems to lack some science.
Yang Qiang: If these players and their opponents and referees have mastered a large amount of data, it is still predictable, but for each team, we have less data, we can only make some speculations. For example, the Brazilian team is very similar to the Argentine team, and the Dutch team may be close to the German team. This speculation may not be accurate in itself. Multiplied by more than ten years of data accumulation, it may be that the German team more than ten years ago and the current German team are also very different. In addition, they have met with different referees and coaches. Maybe their performance is not the same, so you say Science is not enough or because of insufficient data.
Lei Fengwang: You just made a judgment on unsupervised learning on the stage. There are many scientists who are currently devoted to doing research in this area. Did you say that there are actually not many successful cases?
Yang Qiang: Unsupervised learning should be said to be a high goal in academia. However, there are not many successful cases in the industry. However, not many people in the industry do not indicate that it should not be done in the academic world. On the contrary, the academic community should go Doing and attacking what the industry does not do. However, at this stage, we have achieved much better tasks in many areas of deep learning, and there are still few unsupervised cases of success.
Lei Feng Network: I know that you are doing AI emotional research, you introduce your research progress it?
Yang Qiang: We now have some professors who interact with the human-robots. We have a laboratory where we interact with humans. There are some professors who are studying how to identify people’s emotions. For example, computer graphics, speech, and human gestures can be used. It is happy or tired to identify this person, or tired, but also through rules and machine learning methods to let the robot to solve his troubles for human emotions, or to enhance his happiness, we are doing these aspects of the study, and then apply In robots, especially dialogue systems.
Lei Feng network: In fact, AI itself is impossible to have emotions?
Yang Qiang: Yes, it has no emotion in itself. Its emotions are designed by us. So it seems to others that it has emotions, but our designers only have it, we are a mathematical formula for ourselves. Go in.
Lei Feng Network: This emotion recognition technology is now mature?
Yang Qiang: I think this is not mature enough, because there are not enough data collections in this area. There may be some small examples. Successfully displaying emotional and emotional robots on these examples, but I think we still have to wait. Days to collect these data.
Lei Feng network: On the knowledge transfer I do not find most of the information on the Internet, should be more information in English, (non-professional people understand it is difficult), you can briefly explain what it is mainly to solve the problem?
Yang Qiang: It mainly solves two problems. For example, if we open a new shop, we sell a new type of cake. Maybe we don't have any data. In this case, we have no way to recommend users. But if we know that users have a lot of data in other areas, say, drinks, we use this data to build a model. We know that the user's beverage habits and pastry habits may be related, we can The recommendation model of the beverage was successfully transferred to the field of cakes, so that as the number of cakes with the data is small, but some cakes that the user may like can be successfully recommended.
There is already a lot of data in a field, a model can be successfully built, and there is not much data in another field. However, it is related to the previous field and we can migrate that model. The problem that this solves is the problem of less data.
The second problem that can be solved is the problem of personalization. Each of us wants our cell phone to remember some of our habits so that we don’t need to set it every time. How can we make this phone remember this? ? In fact, by migrating and learning, a generic user's mobile phone model can be migrated to personalized data. This will be used more and more later.
Prof. Qiang Yang is the guest speaker of the CCF-GAIR artificial intelligence and robotics summit held in Shenzhen next month. We will leave more information on the research progress and application of transfer learning, reinforcement learning, and emotional recognition. Ticket purchase participants will enjoy a 30% discount; if they can get together for 5 people, they can also choose more discounted 50% group tickets.