Downhill
Reverse mode automatic differentiation in Haskell style.
Overview
Downhill library proposes an approach to automatic differentiation that is both well typed and simple.
Like Backprop, it allows variables to have different types. Traditionally reverse mode automatic differentiation works by constructing computational graph or Wengert list. However, faithfully representing heterogeneous graph in Haskell would result in horrible types.
This library exploits linearity of derivative and constructs linear computational graph to keep everything simple. Bringing heterogeneity to linear graphs don’t make them more complicated. It makes working with them much easier, thanks to parametric polymorphism.
Related work
Backprop
Backprop introduced heterogeneous reverse mode differentiation as a usable library. It has a well typed and nice interface, but it relies on rather complicated machinery under the hood. Downhill library explores alternative ways to implement automatic differentiation.
Conal Elliott’s paper
The Simple Essence of Automatic Differentiation by Conal Elliott explains how gradient is a linear map and all chain rules is just compositions of linear maps. The idea of linear graphs was inspired by this paper.