f-GAN: Training Generative Neural Samplers using Variational Divergence\n Minimization

Abstract

Generative neural samplers are probabilistic models that implement sampling\nusing feedforward neural networks: they take a random input vector and produce\na sample from a probability distribution defined by the network weights. These\nmodels are expressive and allow efficient computation of samples and\nderivatives, but cannot be used for computing likelihoods or for\nmarginalization. The generative-adversarial training method allows to train\nsuch models through the use of an auxiliary discriminative neural network. We\nshow that the generative-adversarial approach is a special case of an existing\nmore general variational divergence estimation approach. We show that any\nf-divergence can be used for training generative neural samplers. We discuss\nthe benefits of various choices of divergence functions on training complexity\nand the quality of the obtained generative models.\n