Variational Autoencoders for Discovering Influential Latent Factors

Abstract

Generative modeling is increasingly being used to simulate or generate new unseen data instances by means of modeling the statistical distribution of data. Generative modeling falls under the broad area of representation learning, which aims to discover representations required for detecting features, classification, and other ways of understanding data. In this vein, variational autoencoders (VAEs) and their variants are one technique of generative modeling (and therefore representation learning) using variational inference under the assumption that the underlying data distribution is composed of a few latent random variables. For example, a VAE (or some other generative learning model) might learn that an image of a person can be generated from the hair color, face shape, and background color. By decomposing the data into latent factors, we could generate and explore new unseen data, which would enable us to investigate how certain data looks like in different environments. However, VAEs are not perfect, and the trained latent factors trained could potentially contain redundant information. In this thesis, we propose to apply VAEs as an unsupervised technique (i.e., in the absence of any external metadata) to investigate the extent to which we can discover a disentangled representation of tabular data and use these factors to generate new data.