A path-norm toolkit for modern networks: consequences, promises and challenges - Archive ouverte HAL Access content directly
Preprints, Working Papers, ... Year : 2023

A path-norm toolkit for modern networks: consequences, promises and challenges

Abstract

This work introduces the first toolkit around path-norms that fully encompasses general DAG ReLU networks with biases, skip connections and any operation based on the extraction of order statistics: max pooling, GroupSort etc. This toolkit notably allows us to establish generalization bounds for modern neural networks that are not only the most widely applicable path-norm based ones, but also recover or beat the sharpest known bounds of this type. These extended path-norms further enjoy the usual benefits of path-norms: ease of computation, invariance under the symmetries of the network, and improved sharpness on layered fully-connected networks compared to the product of operator norms, another complexity measure most commonly used. The versatility of the toolkit and its ease of implementation allow us to challenge the concrete promises of path-norm-based generalization bounds, by numerically evaluating the sharpest known bounds for ResNets on ImageNet.
Fichier principal
Vignette du fichier
Gonon_ICLR_24.pdf (966.93 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04225201 , version 1 (02-10-2023)
hal-04225201 , version 2 (19-10-2023)
hal-04225201 , version 3 (24-11-2023)
hal-04225201 , version 4 (13-03-2024)

Identifiers

  • HAL Id : hal-04225201 , version 4

Cite

Antoine Gonon, Nicolas Brisebarre, Elisa Riccietti, Rémi Gribonval. A path-norm toolkit for modern networks: consequences, promises and challenges. 2023. ⟨hal-04225201v4⟩
109 View
70 Download

Share

Gmail Mastodon Facebook X LinkedIn More