Abstract

We propose a simple, tractable lower bound on the mutual informationcontained in the joint generative density of any latent variable generativemodel: the GILBO (Generative Information Lower BOund). It offers adata-independent measure of the complexity of the learned latent variabledescription, giving the log of the effective description length. It iswell-defined for both VAEs and GANs. We compute the GILBO for 800 GANs and VAEseach trained on four datasets (MNIST, FashionMNIST, CIFAR-10 and CelebA) anddiscuss the results.

 

Quick Read (beta)