Shulou(Shulou.com)11/24 Report--

On September 26, Yann LeCun, chief AI scientist at Facebook parent company Meta, believes that most current AI methods will never bring real intelligence, and he is skeptical of many of the most successful research methods in the field of deep learning today.

The Turing Award winner said that the pursuit of peers is necessary, but it is far from enough. It includes the research of large language models, such as GPT-3 based on Transformer. As Le Kun describes it, supporters of Transformer believe: "AI stands out by tagging everything and training giant models to make discrete predictions."

"they're not wrong," Le Kun explained. In this sense, this may be an important part of the future intelligent system, but I think it lacks the necessary part. " Lecun perfected the use of convolution neural networks, which have achieved incredible results in deep learning programs.

Lecun also saw shortcomings and limitations in many other highly successful areas of the discipline. He insists that intensive learning will never be enough. Researchers like DeepMind's David Silver (David Silver), who have developed the AlphaZero program and mastered chess and go, focus on "very action-oriented" programs, while Le Kun observes that "most of our learning is not done by taking action, but by observation."

Mr Lecun, 62, has a strong urgency to confront a dead end where he thinks many people may be running and try to steer his field in the direction he thinks it should go. "We've seen a lot of talk about what we should do to push AI to the level of human intelligence, and I think some of them are wrong," Lecun said. "our smart machines haven't even reached the level of cat intelligence yet. So why don't we start here?"

Lekun believes that not only academia, but also the AI industry also needs profound reflection. He said it seemed "a little too optimistic" that the self-driving car community, such as Wayve, thought they could "learn almost anything by throwing data at large neural networks."

"you know, I think it's entirely possible for us to have an L5 self-driving car without common sense, but you have to work on design," Lecun said. " He believes that this overdesigned autopilot technology will become as fragile as all computer vision programs that have been eliminated by deep learning. "in the final analysis, there will be a more satisfactory and possibly better solution that involves systems that better understand how the world works," he said. "

"you have to take a step back and say,'OK, we built a ladder, but we want to go to the moon, and this ladder won't take us there," he said, hoping to get people to rethink the basic concept of AI. What I'm trying to say is, it's like making rockets. I can't tell you the details of how we make rockets, but I can provide basic principles. "

Lecun believes that AI systems need to be able to reason, and the process he advocates is to minimize certain potential variables. This enables the system to plan and reason. In addition, Lecun believes that the probability framework should be abandoned, because it is difficult to deal with when we want to do something, such as capturing dependencies between high-dimensional continuous variables. Lekun also advocates abandoning the generation model because the system will have to invest too much resources to predict things that are difficult to predict and may consume too many resources.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.