Using Deep Generative Models for Representation Learning with Applications to AI Explainability

In the last few years, the great potentiality of deep learning has been proven by cutting-edge artificial intelligence applications. Ranging from well-defined tasks like object recognition or speech translation to the more generic autonomous driving, deep learning has been proven to be an extremely powerful tool. One fundamental concept on which deep learning is implicitly built upon is representation learning: how we represent data matters in how we understand the world. In this work, we first study the concept of representation learning, the methods to perform it and its various applications. With the goal of contextualizing this framework to deep generative models, we study the most popular generative methods and make a qualitative comparison between them. In the last section, we will focus our attention on the InfoGAN approach, which imposes explicit conditions on the input representation and is based on the very-well performing Generative Adversarial Network model. Experiments are carried out with InfoGAN on increasingly complex datasets to assess its performance and limitations. We finally discuss how to use GANs in the rising popular field of AI explainability.