Differentiable Programming in High−Energy Physics

Abstract

A key component to the success of deep learning is the use of gradient-based optimization. Deep learning practitioners compose a variety of modules together to build a complex computational pipeline that may depend on millions or billions of parameters. Differentiating such functions is enabled through a computational technique known as automatic differentiation. The success of deep learning has led to an abstraction known as differentiable programming, which is being promoted to a first-class citizen in many programming languages and data analysis frameworks. This often involves replacing some common non-differentiable operations (eg. binning, sorting) with relaxed, differentiable analogues. The result is a system that can be optimized from end-to-end using efficient gradient-based optimization algorithms. A differentiable analysis could be optimized in this way—basic cuts to final fits all taking into account full systematic errors and automatically analyzed. This Snowmass LOI outlines the potential advantages and challenges of adopting a differentiable programming paradigm in high-energy physics.