I think that it was solving the ImageNet challenge. This happened several years ago. This was very closely related to the problem that I was just telling you about, deciding what is in the photo. ImageNet is a database of millions of photographs that have all been labeled by hand with what is in there.

This seems like a very simple task to do. It’s not something that a three-year-old would have any problems with. But it’s something that is absolutely not amenable to the kind of serial computation that has really dominated computing since its birth.

These deep neural nets -- which again, they’re very similar to what’s been there from the very beginning of computation, but trained appropriately and using modern computational power -- are able to say what is in the image.

The fact that they were able to solve this problem that it’s so easy for us to solve, and that they did it in ways that look very much like the way real brains do it, was really very eye-opening to me, and made me feel like we’re at the brink of a revolution with respect to this technology.

Keyboard shortcuts

j previous speech k next speech