Timezone: »

Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples
Samarth Sinha · Zhengli Zhao · Anirudh Goyal · Colin A Raffel · Augustus Odena

Thu Dec 10 09:00 PM -- 11:00 PM (PST) @ Poster Session 6 #1875

We introduce a simple (one line of code) modification to the Generative Adversarial Network (GAN) training algorithm that materially improves results with no increase in computational cost. When updating the generator parameters, we simply zero out the gradient contributions from the elements of the batch that the critic scores as least realistic'. Through experiments on many different GAN variants, we show that thistop-k update' procedure is a generally applicable improvement. In order to understand the nature of the improvement, we conduct extensive analysis on a simple mixture-of-Gaussians dataset and discover several interesting phenomena. Among these is that, when gradient updates are computed using the worst-scoring batch elements, samples can actually be pushed further away from the their nearest mode. We also apply our method to state-of-the-art GAN models including BigGAN and improve state-of-the-art FID for conditional generation on CIFAR-10 from 9.21 to 8.57.

Author Information

Samarth Sinha (University of Toronto, Vector Institute)
Zhengli Zhao (UCI, Google Brain)

Zhengli Zhao is a Computer Science PhD student at UC Irvine, concentrating on deep learning in natural language processing and computer vision.

Anirudh Goyal (Université de Montréal)
Colin A Raffel (Google Brain)

My research focuses on machine learning techniques for sequential data. I am currently a resident at Google Brain. I recently completed a PhD in Electrical Engineering at Columbia University In LabROSA, supervised by Dan Ellis. My thesis focused on learning-based methods for comparing sequences. In 2010, I received a Master's in Music, Science and Technology from Stanford University's CCRMA, supervised by Julius O. Smith III. I did my undergrad at Oberlin College, where I majored in Mathematics.

Augustus Odena (Google Brain)

More from the Same Authors