Deep Probabilistic Generative Models

Master 2 AI (2020-2021).

Recently, generative models have (again) become a hot topic in machine learning thanks to recent advances in deep learning. One of the benefit of these models is their ability to generate new data, see for example:

• face generation: https://thispersondoesnotexist.com/
• word generation: https://thisworddoesnotexist.com/

Moreover, they can be used for semi-supervised learning, feature extraction via latent variables, …

In this course, we will first review the theoretical background required to understand modern generative models:

• Probabilities,
• Latent variables models,
• Expectation Maximization algorithm,
• Change of variable theorem,
• Neural parameterization of probability distributions.

Based on this, we will study modern generative models based on neural networks:

• Variational Auto-Encoders (VAEs),
• Generative Adversarial Networks (GANs),
• Flow models,
• Energy networks.