Kotlin∇: A shape-safe DSL for differentiable programming

Breandan Considine, Michalis Famelis, Liam Paull

Keywords: kotlin, differentiation, type safe

TL;DR: Automatic differentiation in Kotlin with compile-time shape verification for array-programming.

Abstract: Kotlin is a statically-typed programming language with support for embedded domain specific languages, asynchronous programming, and multi-platform compilation. In this work, we present an algebraically-based implementation of automatic differentiation (AD) with shape-safe tensor operations, written in pure Kotlin. Our approach differs from existing AD frameworks in that Kotlin∇ is the first shape-safe AD library fully compatible with the Java type system, requiring no metaprogramming, reflection or compiler intervention to use. A working prototype is available: https://github.com/breandan/kotlingrad.