Revisiting Unbiased Implicit Variational Inference

We develop unbiased implicit variational inference ( ), a method that expands the applicability of variational inference by defining an expressive variational family. considers an implicit variational distribution obtained in a hierarchical manner using a simple reparameterizable distribution whose variational parameters are defined by arbitrarily flexible deep neural networks. Unlike previous works, directly optimizes the evidence lower bound ( ) rather than an approximation to the . We demonstrate on several models, including Bayesian multinomial logistic regression and variational autoencoders, and show that achieves both tighter and better predictive performance than existing approaches at a similar computational cost.