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

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