%0 Conference Paper %1 pmlr-v89-molchanov19a %A Molchanov, Dmitry %A Kharitonov, Valery %A Sobolev, Artem %A Vetrov, Dmitry %B Proceedings of Machine Learning Research %D 2019 %E Chaudhuri, Kamalika %E Sugiyama, Masashi %I PMLR %K optimization variational %P 2593--2602 %T Doubly Semi-Implicit Variational Inference %U http://proceedings.mlr.press/v89/molchanov19a.html %V 89 %X We extend the existing framework of semi-implicit variational inference (SIVI) and introduce doubly semi-implicit variational inference (DSIVI), a way to perform variational inference and learning when both the approximate posterior and the prior distribution are semi-implicit. In other words, DSIVI performs inference in models where the prior and the posterior can be expressed as an intractable infinite mixture of some analytic density with a highly flexible implicit mixing distribution. We provide a sandwich bound on the evidence lower bound (ELBO) objective that can be made arbitrarily tight. Unlike discriminator-based and kernel-based approaches to implicit variational inference, DSIVI optimizes a proper lower bound on ELBO that is asymptotically exact. We evaluate DSIVI on a set of problems that benefit from implicit priors. In particular, we show that DSIVI gives rise to a simple modification of VampPrior, the current state-of-the-art prior for variational autoencoders, which improves its performance.

@inproceedings{pmlr-v89-molchanov19a, abstract = {We extend the existing framework of semi-implicit variational inference (SIVI) and introduce doubly semi-implicit variational inference (DSIVI), a way to perform variational inference and learning when both the approximate posterior and the prior distribution are semi-implicit. In other words, DSIVI performs inference in models where the prior and the posterior can be expressed as an intractable infinite mixture of some analytic density with a highly flexible implicit mixing distribution. We provide a sandwich bound on the evidence lower bound (ELBO) objective that can be made arbitrarily tight. Unlike discriminator-based and kernel-based approaches to implicit variational inference, DSIVI optimizes a proper lower bound on ELBO that is asymptotically exact. We evaluate DSIVI on a set of problems that benefit from implicit priors. In particular, we show that DSIVI gives rise to a simple modification of VampPrior, the current state-of-the-art prior for variational autoencoders, which improves its performance.}, added-at = {2019-12-11T13:58:48.000+0100}, author = {Molchanov, Dmitry and Kharitonov, Valery and Sobolev, Artem and Vetrov, Dmitry}, biburl = {https://www.bibsonomy.org/bibtex/27b44ee4dc29c5293c41d8d4efa9819cd/kirk86}, booktitle = {Proceedings of Machine Learning Research}, description = {Doubly Semi-Implicit Variational Inference}, editor = {Chaudhuri, Kamalika and Sugiyama, Masashi}, interhash = {2ef1ff9b81d3d24e85e79ee09bb4302e}, intrahash = {7b44ee4dc29c5293c41d8d4efa9819cd}, keywords = {optimization variational}, month = {16--18 Apr}, pages = {2593--2602}, pdf = {http://proceedings.mlr.press/v89/molchanov19a/molchanov19a.pdf}, publisher = {PMLR}, series = {Proceedings of Machine Learning Research}, timestamp = {2019-12-11T13:58:48.000+0100}, title = {Doubly Semi-Implicit Variational Inference}, url = {http://proceedings.mlr.press/v89/molchanov19a.html}, volume = 89, year = 2019 }