Higher-Order Automatic Differentiation and Its Applications

Abstract

Differentiable programming is a new paradigm for modeling and optimization in many fields of science and engineering, and automatic differentiation (AD) algorithms are at the heart of differentiable programming. Existing methods to achieve higher-order AD often suffer from one or more of the following problems: (1) exponential scaling with respect to order due to nesting first-order AD; (2) ad-hoc handwritten higher-order rules which are hard to maintain and do not utilize existing first-order AD infrastructures; (3) inefficient data representation and manipulation that causes significant overhead at lowered-order when compared to nesting highly-optimized first-order AD libraries. By combining advanced techniques in computational science, i.e., aggressive type specializing, metaprogramming, and symbolic computing, we introduce a new implementation of Taylor mode automatic differentiation in Julia that addresses these problems. The new implementation shows that it is possible to achieve higher-order AD with minimal overhead and without sacrificing the performance of lower-order AD and obtain significant speedup in real-world scenarios over the existing Julia AD library. In addition, this implementation automatically generates higher-order AD rules from first-order AD rules, which is a step towards a general framework for higher-order AD.