Machine learning, neural networks and deep learning are all buzzwords in AI circles. But essentially they're all referring to the attempt to teach a computer to recognize a dog in any photo, or translate a language in real time, or lots of other tasks.
To do that, engineers present a computer with a lot of data, and teach it by repetition; similar to teaching an animal a trick. They encourage it when it does well, and discourage it when it's wrong. Eventually, the AI will get better at facial recognition or predictive text, or whatever task it's taught.
But the thing is, even if it can spot the dog, the computer has to do the calculation every time. It has to scan every pixel and check everything. It can't anticipate what you ask, it can't realize that usually the dog is here or there in the photo. They don't work that way---until now.
Engineers at 'DeepMind,' Google's AI system, wanted to give their computer a memory. As in, the ability to remember the tasks it had done before and learn from its success and failures on those tasks. And the way they did it is pretty crazy.
We got nominated for a People's Choice Webby! That means, you can help us win. Please take a minute and vote for us here. Thanks!