.. title:: Talk: Fast CHAD

.. only:: html

Fast CHAD
=========

Speaker: Tom Smeding, University of Edinburgh

Abstract
^^^^^^^^

The CHAD algorithm is what you get when you ask a category theorist what AD is.
It enjoys a neat semantical theory with a close relation to the differential
geometry notion of a derivative. But is it fast? It turns out that it is in fact
terribly slow. In this talk we look at the reverse AD version of CHAD and
discuss some of the optimisations one can apply to make the generated gradient
code better (i.e. faster). All our optimisations taken together allow generation
of decent gradient code, although state-of-the-art performance is still future
work.

This talk will (briefly) revisit Efficient CHAD (POPL '24) before moving on to
additional work I did during my PhD.