Apple AI director on artificial intelligence: very interested in reinforcement learning

AsiaIndustrial NetNews: Although artificial intelligence has made great progress in image recognition and product recommendation, this technology still faces many challenges, especially how to make the AI ​​system have a “memory function” is still a difficult problem.

On Tuesday (March 28), Ruslan Salakhutdinov, director of Apple’s AI research division, discussed the limitations of artificial intelligence at the MIT Technology Review conference, but he did not mention how Apple is integrating artificial intelligence into products such as Siri middle.

Salakhutdinov, who joined Apple last October, says he is very interested in reinforcement learning, a technique that teaches computers to iteratively optimize decisions to get the best results. Google, for example, uses reinforcement learning to help data centers achieve optimal cooling and operating configurations to make them more energy efficient.

Researchers at Carnegie Mellon University, where Salakhutdinov is an associate professor, are also doing this recently: using reinforcement learning to train computers to play the 1990s video game “Doom.” Soon, computers learned how to shoot aliens quickly and accurately, and also discovered that dodging could evade enemy fire. However, it is not good at memory, such as the layout of the maze, which makes it impossible to plan and develop game strategies in advance.

Salakhutdinov’s research involves an AI software that records the layout of a virtual maze in “Doom” along with various reference points in order to locate the location of a particular tower. During the game, the software will first determine the color of the torch (red or green), and then locate the corresponding color of the tower according to the color of the torch.

Eventually, the software learned how to find the correct tower in the maze. And when it finds that it has found the wrong tower, it will go back and look for the correct tower. Of particular note is that the software was able to recall the color of the torch each time the tower was found, Salakhutdinov explained.

However, Salakhutdinov said that this type of AI software requires “very long training times” and also requires significant computing power, making it difficult to scale up.

Also, another area Salakhutdinov wants to explore: teaching AI software to learn faster with “less samples and experience.” Although not mentioned in the speech, but his vision is clearly conducive to Apple in a shorter time to create better products.

Some AI experts and analysts believe that Apple’s AI technology is inferior to rivals such as Google and Microsoft due to Apple’s stricter privacy rules that limit the amount of data that can be used to train AI systems. Fortune believes that if Apple had been using less data to train AI systems, it might be able to meet privacy requirements while still being able to make software improvements as quickly as its competitors.

The Links:   3HAC16831-1 3HAC027641-001 IGBT

Published on 09/09/2022