Degeneration in VAE: in the Light of Fisher Information Loss
share
Variational Autoencoder (VAE) is one of the most popular generative models, and enormous advances have been explored in recent years. Due to the increasing complexity of the raw data and the model architecture, deep networks are needed in VAE models while few works discuss their impacts. According to our observation, VAE does not always benefit from deeper architecture: 1) Deeper encoder makes VAE learn more comprehensible latent representations, while results in blurry reconstruction samples; 2) Deeper decoder ensures more high-quality generations, while the latent representations become abstruse; 3) When encoder and decoder both go deeper, abstruse latent representation occurs with blurry reconstruction samples at same time. In this paper, we deduce a Fisher information measure for the corresponding analysis. With such measure, we demonstrate that information loss is ineluctable in feed-forward networks and causes the previous three types of degeneration, especially when the network goes deeper. We also demonstrate that skip connections benefit the preservation of information amount, thus propose a VAE enhanced by skip connections, named SCVAE. In the experiments, SCVAE is shown to mitigate the information loss and to achieve a promising performance in both encoding and decoding tasks. Moreover, SCVAE can be adaptive to other state-of-the-art variants of VAE for further amelioration.
READ FULL TEXT