Deep Probabilistic Programming: TensorFlow Distributions and Edward
The TensorFlow Distribution and Edward libraries implement a vision of probability theory adapted to the modern deep-learning paradigm of end-to-end differentiable computation. We first introduce TensorFlow Distributions, an efficient low-level system for building and manipulating distributions. We focus on the non-obvious design choices in the library, paying particular attention to the Bijector abstraction, which supports composable volume-tracking transformations with automatic caching. We then provide an overview of Edward, a probabilistic programming system built on computational graphs and using Distributions as an efficient backend. In particular, we show how Edward and TensorFlow Distributions can be applied for expanding the frontier of deep generative models and variational inference.
Tue 9 Jan
|14:00 - 15:00|
|15:00 - 15:30|