Just wrote 5 pages for my new paper! I finally have something interesting to present! πΊ I show a geometric interpretation of classifiers. WGAN-GP is different from the Wasserstein Distance and I show what it is actually optimizing. I also talk about RaGANs.
Neat! I just found out why the original Relativistic paired GANs didn't work well except when using a gradient penalty. This paper is going to have lots of interesting tidbits.
I also found gradient penalties which work better than WGAN-GP π and there is a simple theoretical reason to why this is the case. I will release the paper (with code and blog) on Monday or Tuesday next week! πΈ
Ever wondered how to make sense of gradient penalized or 1-Lipschitz discriminators (using spectral normalization) in GANs other than Wasserstein's? My next-week paper will provide an answer and show a framework to derive all these GANs directly.
My new paper is finally coming out tomorrow!!!π
The title is:
"Connections between Support Vector Machines, Wasserstein distance and gradient-penalty GANs"πΈ
Everything is connected!π€―
@gokstudio
I haven't checked conferences yet. The simple enough kind that even I who isn't great at geometry can understand. It's a pretty neat result.