Chalk Talk 400: Machine Learning – cycleGAN

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, 2017, Zhu, Park, Isola, Efros

The Chalk Talk has been away for a while but we’re back with… you guessed it… more machine learning! We’ve talked about Generative Adversarial Networks before. You know, it’s that thing where you lock two neural networks in a room and let them fight it out. If they don’t kill each other, you end up with this cool expert forger and expert detective. (If that doesn’t make any sense, don’t worry, we’re going to go over it again.)

With this paper, they make a tweak – a very cool tweak – that makes the fight more like a choreographed dance. They still want to kill each other, but in a fancy way. Doing it this way gives you more control of the final results. Instead of simply being able to generate a plausible image of a zebra, you can actually turn a photo of a horse into a zebra. And back again! So cool. Especially if you want to be a zebra.

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks – This is the paper we’ll cover.

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks – A nice paper on generative adversarial networks (we covered this before.)