Learning to generate via MMD optimization

Gintare Karolina Dziugaite, Zoubin Ghahramani, Daniel Roy
Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence, PMLR R13:702-711, 2015.

Abstract

We consider learning to generate samples from an unknown distribution given i.i.d. data. In particular, learning is cast as optimizing a transport function to minimize a two-sample test statistic—informally speaking, a good transport function produces samples that cause a two-sample test to fail to reject the null hypothesis. As our objective function, we use an unbiased estimator of the maximum mean discrepancy that is the test statistic underlying the kernel two-sample test proposed by Gretton et al. (2012). We compare to recent proposals for learning generative models and give bounds on the generalization error incurred from optimizing the empirical MMD.

Cite this Paper

BibTeX
@InProceedings{pmlr-vR13-dziugaite15a, title = {Learning to generate via {MMD} optimization}, author = {Dziugaite, Gintare Karolina and Ghahramani, Zoubin and Roy, Daniel}, booktitle = {Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence}, pages = {702--711}, year = {2015}, editor = {Meila, Marina and Heskes, Tom}, volume = {R13}, series = {Proceedings of Machine Learning Research}, month = {12--16 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/r13/main/assets/dziugaite15a/dziugaite15a.pdf}, url = {https://proceedings.mlr.press/r13/dziugaite15a.html}, abstract = {We consider learning to generate samples from an unknown distribution given i.i.d. data. In particular, learning is cast as optimizing a transport function to minimize a two-sample test statistic—informally speaking, a good transport function produces samples that cause a two-sample test to fail to reject the null hypothesis. As our objective function, we use an unbiased estimator of the maximum mean discrepancy that is the test statistic underlying the kernel two-sample test proposed by Gretton et al. (2012). We compare to recent proposals for learning generative models and give bounds on the generalization error incurred from optimizing the empirical MMD.}, note = {Reissued by PMLR on 04 October 2026.} }
Endnote
%0 Conference Paper %T Learning to generate via MMD optimization %A Gintare Karolina Dziugaite %A Zoubin Ghahramani %A Daniel Roy %B Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence %C Proceedings of Machine Learning Research %D 2015 %E Marina Meila %E Tom Heskes %F pmlr-vR13-dziugaite15a %I PMLR %P 702--711 %U https://proceedings.mlr.press/r13/dziugaite15a.html %V R13 %X We consider learning to generate samples from an unknown distribution given i.i.d. data. In particular, learning is cast as optimizing a transport function to minimize a two-sample test statistic—informally speaking, a good transport function produces samples that cause a two-sample test to fail to reject the null hypothesis. As our objective function, we use an unbiased estimator of the maximum mean discrepancy that is the test statistic underlying the kernel two-sample test proposed by Gretton et al. (2012). We compare to recent proposals for learning generative models and give bounds on the generalization error incurred from optimizing the empirical MMD. %Z Reissued by PMLR on 04 October 2026.
APA
Dziugaite, G.K., Ghahramani, Z. & Roy, D.. (2015). Learning to generate via MMD optimization. Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence, in Proceedings of Machine Learning Research R13:702-711 Available from https://proceedings.mlr.press/r13/dziugaite15a.html. Reissued by PMLR on 04 October 2026.

Related Material