Document de travail

A mathematical model for automatic differentiation in machine learning

Jérôme Bolte et Edouard Pauwels

Résumé

Automatic differentiation, as implemented today, does not have a simple mathematical model adapted to the needs of modern machine learning. In this work we articulate the relationships between differentiation of programs as implemented in practice and differentiation of nonsmooth functions. To this end we provide a simple class of functions, a nonsmooth calculus, and show how they apply to stochastic approximation methods. We also evidence the issue of artificial critical points created by algorithmic differentiation and show how usual methods avoid these points with probability one.

Remplacé par

Jérôme Bolte et Edouard Pauwels, « A mathematical model for automatic differentiation in machine learning », dans Advances in Neural Information Processing Systems, sous la direction de Hugo Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan et H. Lin, vol. 33, 2020.

Référence

Jérôme Bolte et Edouard Pauwels, « A mathematical model for automatic differentiation in machine learning », TSE Working Paper, n° 21-1184, janvier 2021.

Voir aussi

Publié dans

TSE Working Paper, n° 21-1184, janvier 2021