AI learns to beat your high scores just by watching games

While pinball and boxing games were easiest, the AI surprised researchers with its capacity to learn strategies

Google bought AI company DeepMind last year and what has the latter been doing? Creating AI that can beat gamers on the Atari 2600, apparently. The software was taught 49 different games on the platform and bested human top scores on 23 of them. Pretty terrifiying if you're a pro gamer.

DeepMind co-founder Demis Hassabis claims the algorithm behind the software is the first that can match human performance across "a wide range of challenging tasks". 

First games, then the world?

The algorithm Hassabis speaks of is also called a "deep neural network". They are often used for image recognition, but DeepMind was clever enough to add what is called reinforcement learning. That's when you're rewarded for doing something right, such as being awarded top scores for playing a game well.

Teaching AI games isn't anything new. What is new is the amount of computing power available to us that wasn't in the early days of AI. Watching an Atari game would involve processing about 2 million pixels of data a second and being able to harness that kind of data churning could be useful to the company that has access to data from anyone with a Gmail account.

It's likely Google is already using DeepMind tech in its products, most likely in advertising and figuring out what ads create the best results. Or it could help with Google's autonomous cars.

We're hoping that this won't eventually translate into tech where bots whoop our ass in Call of Duty.

[Source: Newscientist]