Wasserstein Variational Inference

About

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.

Luca Ambrogioni, Umut Güçlü, Yağmur Güçlütürk, Max Hinne, Eric Maris, Marcel A. J. van Gerven · 2018

Related benchmarks

Task	Dataset	Result
Few-shot Image Classification	tieredImageNet	Accuracy0.871	190
Few-shot classification	MiniImagenet	5-way 5-shot Accuracy—	98
Few-shot classification	FC-100	Accuracy68.5	36
Image Classification	MEDIC subset	AA59.8	36
Few-shot Image Classification	AIDER subset simulation	Average Accuracy (AA)66.1	36
Few-shot classification	CIFAR FS (test)	5-way 1-shot Acc77.9	26
Few-shot classification	CDD 5-way 1-shot	AA61.2	19
Few-shot classification	CDD 5-way 5-shot	Average Accuracy (AA)65.8	17