In the just-concluded CCAI, today's headline scientist Li Lei shared with us some of his thoughts on the future direction of machine learning, and discussed on this basis what we need to do to get closer to general purpose AI. problem. Today's headline was originally a news aggregation platform, but there is a tradition of applying AI to solve problems. Recently, today's headline has also made its own news robot and has completed a large number of reports during the Rio Olympic Games. Li Lei is a scientist with many years of research experience in machine learning, artificial intelligence and deep learning. It also has its own opinion on this issue.
Li Lei said in his speech that to figure out how far we are from the general robot, we must first understand three aspects: First, what is artificial intelligence? The second is how much we can do now? After clarifying these two issues, we can study the third aspect: Where are our limitations? Where are our challenges? How do we go to tomorrow?
In fact, not only Li Lei, but also many guests have mentioned in the speeches and discussions in the conference: We need to manage ourselves and consumers' perception of artificial intelligence. At present, people’s expectations for artificial intelligence are generally high. If we do not have the means to raise the actual level of artificial intelligence to such a high level in a short period of time (actually, we are currently virtually impossible to do this), it may be too high. The disillusionment of expectations will lead to a new round of depression. Therefore, we need to remind ourselves of what kind of state the current artificial intelligence is.
What is artificial intelligence?
What is artificial intelligence? Li Lei said that artificial intelligence actually has two definitions. One type of artificial intelligence is to make machines think like people, to make decisions, to solve problems, to have the ability to learn, and to have the ability to act. In short, it is all people-oriented.
There is also a definition called rational intelligence. The goal of rational intelligence is not to compare machines to people, but to think of computing as a natural phenomenon, to study what kind of reasonable levels this natural phenomenon can achieve in the realm of intelligence, and how to do it. .
What is the degree of artificial intelligence?
The research content of artificial intelligence is very wide, and machine learning is just one of them. On specific issues, artificial intelligence has reached or even surpassed human standards. Li Lei gave a lot of examples in this regard, including AlphaGo, which is well-known to all, robots and pictures, as well as the Olympic news robot recently developed on the headline today. More and more work done by people before can already be done with robots. In these specific areas, the capabilities of robots have reached or even surpassed humanity.
Where are our limitations and challenges?
However, Li Lei said that we are still far away from general-purpose artificial intelligence. He said that in the study, they found that the problems that deep learning is currently best at are those that have “supervised learningâ€. Many of the examples he cited before depended on supervised learning.
However, Li Lei said that we must know that artificial intelligence, or machine learning, is not only supervised learning, but also deep learning. The problems that artificial intelligence needs to solve are actually far more than this. And although current deep learning has performed well, it actually has many limitations. For example, it relies on a large amount of annotated data. In the past, the price paid for so much annotated data was actually very high. Even today, many people don't have the conditions to acquire so much data.
Another limitation of deep learning is that its versatility is not strong enough. The program for playing chess will only play chess. The program for identifying cats and dogs will only recognize cats and dogs. It is almost impossible for you to let them do other things.
What efforts should we make?
Li Lei said that we can try to solve, in order to promote the development of machine learning, there are three issues:
First, we need to develop its interpretability, which can also be said to be the theoretical basis. When machine learning is good or bad, successful or unsuccessful, we need to understand in principle why it has such a performance. So that we can fundamentally improve it.
The second is: I hope machine learning can do more reasoning than simple judgment.
The third is the past research, found that machine learning requires a lot of computing resources to train, use a lot of GPUs and computers to train, at the same time it also consumes a lot of energy, Li Lei said he hopes to find a The method reduces the energy consumption and can achieve good results without the need for such exaggerated computing power and energy consumption. If we can achieve these three points, I believe we can go further in the study of machine learning.
Fpc Connector,Straight Pin Fpc Connector,Gold-Plated Straight Pin Fpc Connector,Conventional Fpc Connector
Shenzhen Jinyicheng Electronci Technology Co.,Ltd. , https://www.jycconnector.com