Backpropagation in the Simply Typed Lambda-calculus with Linear Negation
This provides a purely logical understanding of backpropagation for differentiable programming, enabling more synthetic and modular expression of computational graphs in fields like machine learning.
The paper tackles the problem of extending backpropagation to higher-order programming languages by defining a compositional transformation from the simply-typed lambda-calculus to itself with linear negation, proving it computes gradients as efficiently as first-order backpropagation.
Backpropagation is a classic automatic differentiation algorithm computing the gradient of functions specified by a certain class of simple, first-order programs, called computational graphs. It is a fundamental tool in several fields, most notably machine learning, where it is the key for efficiently training (deep) neural networks. Recent years have witnessed the quick growth of a research field called differentiable programming, the aim of which is to express computational graphs more synthetically and modularly by resorting to actual programming languages endowed with control flow operators and higher-order combinators, such as map and fold. In this paper, we extend the backpropagation algorithm to a paradigmatic example of such a programming language: we define a compositional program transformation from the simply-typed lambda-calculus to itself augmented with a notion of linear negation, and prove that this computes the gradient of the source program with the same efficiency as first-order backpropagation. The transformation is completely effect-free and thus provides a purely logical understanding of the dynamics of backpropagation.