One of the uses of Conal Elliott’s ConCat infrastructure for compiling with categories is implementing Deep Learning. ConCat implements the automatic differentiation needed to compute gradients. It can also be used to transform the resulting code to run on a GPU via the Accelerate framework. Composing both yields high-performance GPU-based Deep Learning, implemented as plain Haskell code. The talk describes how these pieces fit together in an industrial application, what the advantages are over the dominant Python-based frameworks, as well as challenges of composing ConCat and Accelerate.

Mike Sperber is CEO of Active Group GmbH in Tübingen, Germany.