Parallel Algorithmic Differentiation for OCaml

Markus Mottl presents AD-OCaml, a framework for algorithmic differentiation in OCaml supporting nested derivatives and parallel execution.

Markus Mottl C◦mp◦se :: Conference composeconference.org/ February 5, 2016 Slides: github.com/cloudRoutine/composeconf2016 AD-OCaml is a framework for algorithmic differentiation of OCaml programs with the following features: support for all language constructs; forward and reverse mode AD including arbitrary nesting of derivatives; fast vector and matrix operations; support for imperative operations in the presence of aliasing; implicit data- and task-parallel execution; visualization of data flow graphs; scalability and extensibility. The talk will provide background information on algorithmic differentiation, the challenges of implementing the described features, and its applications in the field of machine learning and optimization. The talk may include a demonstration of an alpha version of the framework, which, as of now, already supports the aforementioned features.