Gigaom AI Minute – October 18

Transcript

While, many of the techniques that we use in artificial intelligence are actually pretty old, every now and then a new idea comes along. One example is generative adversarial networks, called GANs, that promise to make advances in unsupervised learning.

They go like this, say the object is to figure out an algorithm to identify if a photo has a cat in it. You have one computer, the generative one, which is trying to say, this is a cat and this isn't, and then you have another completely unrelated system, the adversarial one, who's trying to tell the difference between reality and the solutions offered by the generative one.

The hope is that at some point it gets to the point that the adversarial model can't tell the difference between reality and what the generative model is producing. In other words it's almost as if the generative model is passing its own kind of Turing test.