April 21st 2021, 11:00 (zoom link)

Abstract. Automatic differentiation (AD) has a split personality. While scientific applications of AD have had decades to mature, the same techniques were rediscovered more recently in machine learning, which has built its own parallel universe of tooling with very different design considerations and constraints. Can these two worlds be combined?

We present Zygote, an AD system which attempts to meet the needs of both the ML and scientific communities in one tool, building on work by the AD community. We’ll assess Zygote’s current design and implementation, future challenges to building general-purpose AD, and some of the exciting applications enabled by the merging of ML with traditional scientific modelling.