Adjoints by Automatic Differentiation

Abstract : We present Automatic Differentiation (AD), a technique to obtain derivatives of functions provided as programs. We present the principles that justify why AD is possible and explain its performance. The adjoint mode of AD is the choice approach to obtain gradients, like the gradients needed for data assimilation. We show the specific difficulties of the adjoint mode, and list a few AD tools that handle these problems well. We show why AD needs an enlightened user to achieve optimal efficiency.
Complete list of metadatas

Cited literature [11 references]  Display  Hide  Download

https://hal.inria.fr/hal-01109881
Contributor : Laurent Hascoet <>
Submitted on : Tuesday, January 27, 2015 - 10:41:37 AM
Last modification on : Thursday, January 11, 2018 - 4:20:45 PM
Long-term archiving on : Tuesday, April 28, 2015 - 10:27:38 AM

File

LesHouches.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01109881, version 1

Collections

Citation

Laurent Hascoët. Adjoints by Automatic Differentiation. Advanced Data Assimilation for Geosciences, Oxford University Press, 2014, 978-0-19-872384-4. ⟨hal-01109881⟩

Share

Metrics

Record views

368

Files downloads

393