Deep Generative Models: Pitfalls and Fixes

Abstract

The promise of deep learning is to discover rich, hierarchical models that represent probability distributions over the kinds of data encountered in artificial intelligence applications, such as natural images, audio waveforms containing speech, and symbols in natural language corpora. In recent years, the most striking successes in deep learning have involved generative models. However, in their vanilla forms, generative models have a number of shortcomings and failure modes that can hinder their application: they can be difficult to train on high dimensional data, and they can fail in tasks such as the generation of realistic artificial data. In this thesis, we first explore the reasons for these failures in the adversarial-based generative models and propose a novel approach to alleviate these shortfalls. Then, we discuss how a learned generative model can be employed for a downstream task such as speech recognition.