Normalizing Flows Image Generation using an Invertible Deep Neural Network

Abstract

Normalizing ows are a statistical machine learning method whereby we map a simple Gaussian base probability density distribution to a complex unknown prob- ability density distribution via a series of invertible di erentiable di eomorphism transformations. These transformations are implemented as a deep invertible neu- ral network. This allows us to perform probability density evaluation and sampling which would otherwise be intractable due to complex integrals. In this project we are researching normalizing ows and implementing a version of the recently published Glow architecture to perform image generation, trained using the MNIST dataset, FashionMNIST and the CelebA dataset, to gain an understanding of how normaliz- ing ows work, evaluate the performance and study the practical applications and limitations of this statistical machine learning technique.