Science and Technology News May 25, the end of the second round of the man-machine war, Ko Jie and Alpha Go in the game, there were mistakes, and finally 155 hands in the plate to admit negative. In this regard, Cheetah Mobile CEO Fu Sheng believes that AlphaGo 2.0 has not made a substantive breakthrough, "starting from scratch" AI revolution has a long way to go. Fu Sheng believes that Deepmind, as the world's top in-depth learning institution, leads the exploration of human in-depth learning. But in a year, AlphaGo 2.0 essentially optimizes the algorithm and improves the computing power. It also reminds AI practitioners not only to hope for explosive technological breakthroughs, but also to focus on the combination of AI and application.
Pre-game Cheetah Mobile CEO Fu Sheng predicts that Alpha Go's current version has not been learnt without supervision. If it can be achieved, the new version of technology breakthrough is no less meaningful than the first version of Alpha Go. Training from scratch means using reinforcement learning to evolve from scratch, relying solely on the final Reward of the game. "It is an open question to use RL model to converge to near optimal solution without initial supervision, even for some games of red and white machine. If AlphaGo 2 can complete learning from scratch, it probably means that there is a significant breakthrough in the reinforcement learning algorithm itself. This breakthrough may not only be used in Go, but also have a lot of possibilities for other applications, so its significance will be no less than that of AlphaGo 1.
Fu Sheng's point of view was certified after the game. Aja Huang (Huang Shijie), one of the core authors of Alpha Gos, declared after the first war that "Alpha Gos is a stand-alone version, but there is still training in human knowledge."
"AlphaGo 2.0 did not break through in essence, and the unsupervised learning we were looking forward to did not come. Deepmind, as the world's top in-depth learning institution, leads the exploration of human in-depth learning. But in a year, AlphaGo 2.0 essentially optimizes the algorithm and improves the computing power. It also reminds AI practitioners not only to hope for explosive technological breakthroughs, but also to focus on the combination of AI and application, "Fu Sheng said.
This theory of AI application integration has a long history. Fu Sheng has publicly pointed out that the opportunity of in-depth learning lies in the collection of applications rather than just the output of technology. He believes that deep learning is an algorithmic revolution that essentially reduces technical barriers. Because of the immobilization of the basic algorithm model, the driving force of the algorithm has been greatly reduced, and the algorithm driving has become data driving. Therefore, the core of in-depth learning is data-driven. Although the organizations with model parameters have their own advantages, more data parameters will quickly leveling the advantages.
Essentially, although AI is a technology and tool, due to the rapid development of the Internet, it is very difficult for an independent technology itself to become a complete industry today. It is difficult to imagine a company can succeed through technology export. In the future, deep learning is the basic technology application. Many companies will have the ability of deep learning research and development. All companies will be technology companies. Science and technology are the foundation, and need to be combined with application.
This argument is not empty. Cheetah bought a French news product News Republic last year, turning the user's click behavior into the annotation part of the data, and the neural network will find the relevant news to push automatically. Cheetah also makes Live. me, the largest third-party live broadcasting platform in the United States. Hundreds of thousands of American users broadcast it every day, generating millions and tens of millions of standard faces, which enables cheetahs to find accurate data. Cheetah's face recognition technology is widely used in pornography and child recognition. In the recent evaluation of LFW face recognition, Cheetah has achieved the top three results.
From Deep Blue to Alpha Dog, chess and card games have been used to test the gap between human beings and AI, tracing back the reasons. On the one hand, chess has a long history, human beings have enough accumulation, Go has enough changes, violent search can not solve the problem, AI must have "intuition"; on the other hand, it belongs to Complete Information Game, in fact. The most convenient way to verify AI capabilities. Fu Sheng predicts that AI will play an assistant role in the future, not a confrontational role. It will be an era of coexistence of human and computer.
