Yann LeCun Profile
It is much more difficult for an artificial intelligence system to recognize video content than to identify a picture. Because video contains a lot of information, it is always an unsolved problem in the field of artificial intelligence to understand the rich content of video in the machine.
However, Yann LeCun, head of the Facebook Artificial Intelligence Lab, believes that since humans have taught machines to identify images and even face recognition, the machine can recognize the video. The method of teaching machine learning videos is similar to that of infant learning. That is, let the machine watch the video like a baby and tell it what the video says.
On March 9th, local time, MIT Technology Review published an article about Yan Xun LeCun, head of Facebook's artificial intelligence lab, about machine vision technology. In Yang Lekun's view, there are still many shortcomings in machine vision. I can only understand the knowledge that humans teach it, but this will change in the future: through training, you can show the machine a few frames in a video, it can predict What happens next.
Born in France, Yang Lekun studied post-doctoral research at the University of Toronto with Geoffrey Hinton, the founder of deep learning. He is now a tenured professor at New York University and a pioneer in artificial intelligence neural network research. Prior to joining Facebook, he worked in Bell Labs for more than 20 years, where he was the world's most famous computer research laboratory and produced many great products. Yang Lekun developed a system that recognizes handwritten digits during his work at Bell Labs and named it LeNet. This system automatically recognizes bank checks.
The following is Yang Lekun's understanding of artificial intelligence machine vision:
What progress has been made in machine vision?
There is a main body in a picture. The rule is to let the machine classify various subjects. As long as you have enough data, like 1000 subjects per directory, the machine can be aware of specific categories, such as specific brand cars, specific types of plants or specific breeds of dogs. We can also get to know more abstract categories such as weather maps, landscapes, sunsets, weddings or birthday parties. Just five years ago, we were not quite sure whether the machine had completely solved the problem, but now it does not mean that the machine vision problem has been solved.
What are the important issues that machine vision is not yet "solving"?
People had an idea a few years ago - generating a tag or description for pictures and videos. On the surface, there has been significant progress, but in fact these results are not as significant as they seem. The knowledge of the machine in a particular field is limited to what we teach them. Most of these systems will see the following situations. You can show them some other categories of pictures, or show them some scenes you have never seen before, and the machine will say a bunch of garbage. They have no common sense for the time being.
What is the connection between machine vision and common sense?
It depends on who you are discussing this issue, even if there are different answers within Facebook. You can interact with the smart system in language. But the problem is that language is a low-bandwidth information transmission channel. Information is expressed in words because humans have a lot of background knowledge to understand this information.
Some people think that the only way to provide enough information to the AI ​​system is to use visual perception as the basis, which is higher than the amount of information brought by the language as input. If you tell the machine "This is a smart phone"; "This is a road roller"; "These things can be pushed by you, others can't be moved", maybe the machine can master some basic knowledge of the world like a baby.
Because babies can learn a lot from the world without specific guidance. One of the things we really want to do is let the machine learn by watching videos or through other channels to show them the connections of everything in the real world. This method ultimately allows the machine to master common sense. This is also the way animals and babies learn in the first few months of life – you can learn quite a bit by simply observing the world. There are still many ways to "fool" machines because they have very limited knowledge of the world.
What progress has been made in "Let the machine learn from observation?"
There is a view that machine learning systems should be able to predict the future, and we are very interested in this idea. You show the machine a few frames of a video, and it predicts what will happen next. If we can train the system to do this, I think we have developed a technology based on an unsupervised learning system. I think this is where more interesting things are likely to happen. This kind of application is not necessary in machine vision, but it is indeed an advancement in the field of artificial intelligence.
Smart Bms,Bms For Battery,Bms For Lithium Battery,Bms Module
HuiZhou Superpower Technology Co.,Ltd. , https://www.spchargers.com