Understanding Variational Autoencoders and Disentanglement Metrics

Abstract

In this thesis, we conduct a thorough study of "Variational Autoencoders". We explain the limitations of ``supervised learning" and emphasize the need for ``generative models" to solve complex problems. Variational Autoencoder (VAE) is an effective tool for generative modeling. We understand the rich mathematical basis of VAEs, the evidence lower bound (ELBO) and how regularization of original VAEs brings the trade-off between reconstruction fidelity and quality of disentanglement within the learnt representations. We validate the theory by conducting experiments on Frey's Faces image dataset using an encoder and decoder architecture using convolutional neural networks . In addition to this, we also learn what constitutes a ``good representation" and how "disentanglement metrics" helps in comparing representations obtained from two different models. We briefly describe two different types of disentanglement metrics Beta-VAE metric and Mutual Information Gap (MIG).