Wasserstein Variational Inference
This work addresses the problem of improving variational inference for researchers and practitioners in machine learning by offering a novel framework that enhances stability and applicability, though it appears incremental as it builds on existing variational autoencoding techniques.
The paper tackles approximate Bayesian inference by introducing Wasserstein variational inference, a new method based on optimal transport theory that includes f-divergences and Wasserstein distance as special cases, resulting in a stable likelihood-free training technique applicable to implicit distributions and probabilistic programs, with tests showing robustness and performance improvements in new autoencoder forms compared to existing variational autoencoding methods.
This paper introduces Wasserstein variational inference, a new form of approximate Bayesian inference based on optimal transport theory. Wasserstein variational inference uses a new family of divergences that includes both f-divergences and the Wasserstein distance as special cases. The gradients of the Wasserstein variational loss are obtained by backpropagating through the Sinkhorn iterations. This technique results in a very stable likelihood-free training method that can be used with implicit distributions and probabilistic programs. Using the Wasserstein variational inference framework, we introduce several new forms of autoencoders and test their robustness and performance against existing variational autoencoding techniques.