Getting artificial intelligence systems to remember things like humans do is a big challenge, according to Apple’s director of AI research, Ruslan Salakhutdinov, speaking at an MIT Technology Review conference this week, Fortune reports.
Salakhutdinov is testing a type of AI dubbed reinforcement learning to help solve the problem. Reinforcement learning teaches computers to try different approaches until ultimately reaching the correct or best result. But Salakhutdinov said this type of AI software takes "a long time to train" and requires huge amounts of computing power.
Salakhutdinov has had some success, however, in training computers to play "Doom," a video game from the 1990s. Using reinforcement learning, computers learned the layouts of virtual mazes and eventually learned to navigate the maze to reach the correct tower.
IBM, Intel, Apple, Google and others are all racing to take the lead in AI. But one challenge they all face? A lot of data is required in order to "train" AI systems.
Salakhutdinov said in addition to reinforcement training he is also working on a system that could help AI software learn faster from "few examples and few experiences."
Ultimately, the company that has access to the most data for training their AI systems may come out ahead. IBM, with its long lead time and access to huge troves of data, has real potential to start monetizing enterprise and consumer AI-based systems first.
Smaller companies, or those like Apple that have strict requirements around data use, may have a harder time. But if Salakhutdinov can perfect his system for teaching AI systems with less data, it could change the current scenario completely.