Chi-square Generative Adversarial Network

Chenyang Tao, Liqun Chen, Ricardo Henao, Jianfeng Feng, Lawrence Carin Duke
Proceedings of the 35th International Conference on Machine Learning, PMLR 80:4887-4896, 2018.

Abstract

To assess the difference between real and synthetic data, Generative Adversarial Networks (GANs) are trained using a distribution discrepancy measure. Three widely employed measures are information-theoretic divergences, integral probability metrics, and Hilbert space discrepancy metrics. We elucidate the theoretical connections between these three popular GAN training criteria and propose a novel procedure, called $\chi^2$ (Chi-square) GAN, that is conceptually simple, stable at training and resistant to mode collapse. Our procedure naturally generalizes to address the problem of simultaneous matching of multiple distributions. Further, we propose a resampling strategy that significantly improves sample quality, by repurposing the trained critic function via an importance weighting mechanism. Experiments show that the proposed procedure improves stability and convergence, and yields state-of-art results on a wide range of generative modeling tasks.

Cite this Paper


BibTeX
@InProceedings{pmlr-v80-tao18b, title = {{C}hi-square Generative Adversarial Network}, author = {Tao, Chenyang and Chen, Liqun and Henao, Ricardo and Feng, Jianfeng and Duke, Lawrence Carin}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {4887--4896}, year = {2018}, editor = {Dy, Jennifer and Krause, Andreas}, volume = {80}, series = {Proceedings of Machine Learning Research}, month = {10--15 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v80/tao18b/tao18b.pdf}, url = {https://proceedings.mlr.press/v80/tao18b.html}, abstract = {To assess the difference between real and synthetic data, Generative Adversarial Networks (GANs) are trained using a distribution discrepancy measure. Three widely employed measures are information-theoretic divergences, integral probability metrics, and Hilbert space discrepancy metrics. We elucidate the theoretical connections between these three popular GAN training criteria and propose a novel procedure, called $\chi^2$ (Chi-square) GAN, that is conceptually simple, stable at training and resistant to mode collapse. Our procedure naturally generalizes to address the problem of simultaneous matching of multiple distributions. Further, we propose a resampling strategy that significantly improves sample quality, by repurposing the trained critic function via an importance weighting mechanism. Experiments show that the proposed procedure improves stability and convergence, and yields state-of-art results on a wide range of generative modeling tasks.} }
Endnote
%0 Conference Paper %T Chi-square Generative Adversarial Network %A Chenyang Tao %A Liqun Chen %A Ricardo Henao %A Jianfeng Feng %A Lawrence Carin Duke %B Proceedings of the 35th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2018 %E Jennifer Dy %E Andreas Krause %F pmlr-v80-tao18b %I PMLR %P 4887--4896 %U https://proceedings.mlr.press/v80/tao18b.html %V 80 %X To assess the difference between real and synthetic data, Generative Adversarial Networks (GANs) are trained using a distribution discrepancy measure. Three widely employed measures are information-theoretic divergences, integral probability metrics, and Hilbert space discrepancy metrics. We elucidate the theoretical connections between these three popular GAN training criteria and propose a novel procedure, called $\chi^2$ (Chi-square) GAN, that is conceptually simple, stable at training and resistant to mode collapse. Our procedure naturally generalizes to address the problem of simultaneous matching of multiple distributions. Further, we propose a resampling strategy that significantly improves sample quality, by repurposing the trained critic function via an importance weighting mechanism. Experiments show that the proposed procedure improves stability and convergence, and yields state-of-art results on a wide range of generative modeling tasks.
APA
Tao, C., Chen, L., Henao, R., Feng, J. & Duke, L.C.. (2018). Chi-square Generative Adversarial Network. Proceedings of the 35th International Conference on Machine Learning, in Proceedings of Machine Learning Research 80:4887-4896 Available from https://proceedings.mlr.press/v80/tao18b.html.

Related Material