There are a lot of things that we do without having to think too much about. For example if you’re a chef, there’s a good chance you won’t have to think too hard when it comes to recreating a dish that you’ve done the past 1,000 times. However at the same time, when we pass that knowledge down to someone else, we can recall the steps and break it down.
Unfortunately when it comes to AI, it isn’t always so intuitive, but that has changed as Alphabet’s DeepMind has gained the ability to retain information that it had previously gained and use it again, thanks to the efforts of the DeepMind team and researchers at the Imperial College London.
Speaking to WIRED, James Kirkpatrick, a research scientist at DeepMind says, “Previously, we had a system that could learn to play any game, but it could only learn to play one game. Here we are demonstrating a system that can learn to play several games one after the other.” Kirkpatrick also notes that the inability for neural networks and AI to transfer knowledge is one of its shortcomings, which we can only assume they’re addressing with this latest effort.
In terms of “continual learning”, we have seen it briefly demonstrated in voice assistant software, where users can ask questions and follow-up questions without having to repeat themselves or start from scratch. Obviously that’s just scratching the surface, but wouldn’t it be nice if one day we could have conversations that flow with our robot oppressors?