# Introduction

## What is a Normalizing Flow?​

Normalizing Flows are a family a methods for representing and learning high-dimensional probability distributions. They have found state-of-the-art applications in modeling complex distributions over images, speech, syntactic structure, and molecules, to name a few. Simply put, a Normalizing Flow is a composition of learnable functions that inputs samples from a simple random distribution, typically Gaussian noise, and outputs samples from a more complex target distribution. Here is an illustration (taken with permission from here):

A simple source of noise, $z_0$, is passed through a number of invertible functions, $f_1,f_2,\ldots,f_k$ to produce a more complex random variable, $z_k$. The invertible functions are constructed in a clever way so that we can easily sample from $z_k$ and calculate its density function, $p_k(\cdot)$. The field of Normalizing Flows can be seen as a modern take on the change of variables method for random distributions, where the transformations are high-dimensional, often employing neural networks, and are designed for effective stochastic optimization.

We believe, although still a nascent field, that Normalizing Flows are a fundamental component of the modern Bayesian statistics and probabilistic computing toolkit, and we will likely see many more exciting applications in the near future.

## What is FlowTorch?​

FlowTorch is a library that provides PyTorch components for constructing Normalizing Flows using the latest research in the field. It builds on an earlier sub-library of code from Pyro developed by the author since 2018. The main goals behind creating a new library for Normalizing Flows are to:

• define an elegant interface for Normalizing Flow methodology, building on our experience with Pyro, so that practitioners can easily utilize these methods and researchers can easily contribute their own implementations;
• develop robust unit tests and other code quality practices to guarantee production quality code;
• promote the methods in applied settings by fostering a community of Normalizing Flow practioners and linking them with researchers;
• accelerate research in Normalizing Flows by providing standard implementations, benchmarking, and a comprehensive literature survey.

## Where to From Here?​

We recommend reading the next two sections to install FlowTorch and train your first Normalizing Flow. For more theoretical background on Normalizing Flows and information about their applications, see the primer here and the list of survey papers here.