Jean-Luc Bouchot
1, @, Sri Hari Krishna Narayanan, Narayanan
2
, Paul Hovland
2
1 : Centre Inria de Saclay Institut National de Recherche en Informatique et en Automatique
2 : Argonne National Laboratory [Lemont]
Computing derivatives has been needed since the dawn of (calculus) times as soon as optimisation problems have been of interest. As the base computations complexifies so do the computations of the derivatives up to a point where it becomes unbearable for large pieces of scientific codes. Consequently algiorithmic differentiation has been ubiquitous to scientific computing since its first official developments in the mid 60s. Today, AD is everywhere, and has regained interest with the concepts of backpropagation (a.k.a. adjoint algorithmic differentiation and a corner stone of AI) or differentiable programming. In the right hands, AD facilitates the implementation of gradients / Jacobian / derivatives and thereby help big scientific applications. This hands on tutorial gives you the theoretical basics for understanding what AD is and is not and how it can be used via various frameworks. We will also review performance evaluations and potential improvements in the derivative computations.
Type :
:
Présentation
Thématiques
:
La programmation d’aujourd’hui et de demain
