Normalizing Flows: Theory, Practice and Applications

Normalizing Flow is a technique in machine learning that uses a series of invertible transformations to map a simple probability distribution into a more complex distribution. These transformations allow you to model complex data efficiently while preserving the ability to calculate accurate probability densities. In this seminar we will cover the theory behind flows, in addition to their advantages in the class of generative models. We will explore the most used methods, placing emphasis on the Affine Coupling technique. Finally, we will examine some of its applications, especially in anomaly detection and prediction of stochastic events.