$\partial$ is for Dialectica
Résumé
Automatic Differentiation is the study of the efficient computation of differentials. While the first automatic differentiation algorithms are concomitant with the birth of computer science, the specific backpropagation algorithm has been brought to a modern light by its application to neural networks. This work unveils a surprising connection between backpropagation and Gödel's Dialectica interpretation, a logical translation that realizes semi-classical axioms. This unexpected correspondence is exploited through different logical settings. In particular, we show that the computational interpretation of Dialectica translates to the differential λ-calculus and that Differential Linear Logic subsumes the logical interpretation of Dialectica.
Domaines
Informatique [cs]Origine | Fichiers produits par l'(les) auteur(s) |
---|