Non-convex SGD and Lojasiewicz-type conditions for deep learning | Vidéo | Carmin.tv

00:00:00 / 00:00:00

Non-convex SGD and Lojasiewicz-type conditions for deep learning

De Kevin Scaman

Apparaît dans la collection : Learning and Optimization in Luminy - LOL2022 / Apprentissage et Optimisation à Luminy - LOL2022

First-order non-convex optimization is at the heart of neural networks training. Recent analyses showed that the Polyak-Lojasiewicz condition is particularly well-suited to analyze the convergence of the training error for these architectures. In this short presentation, I will propose extensions of this condition that allows for more flexibility and application scenarios, and show how stochastic gradient descent converges under these conditions. Then, I will show how to use these conditions to prove the convergence of the test error for simple deep learning architectures in an online setting.

Informations sur la vidéo

Date de captation 04/10/2022
Date de publication 10/11/2022
Institut CIRM
Licence CC BY NC ND
Langue Anglais
Audience Chercheurs, Doctorants
Réalisateur(s) Guillaume Hennenfent
Format MP4

Données de citation

DOI 10.24350/CIRM.V.19965303
Citer cette vidéo Scaman, Kevin (04/10/2022). Non-convex SGD and Lojasiewicz-type conditions for deep learning. CIRM. Audiovisual resource. DOI: 10.24350/CIRM.V.19965303
URL https://dx.doi.org/10.24350/CIRM.V.19965303

Domaine(s)

Apprentissage

Bibliographie

SCAMAN, Kevin, MALHERBE, Cédric, et DOS SANTOS, Ludovic. Convergence Rates of Non-Convex Stochastic Gradient Descent Under a Generic Lojasiewicz Condition and Local Smoothness. In : International Conference on Machine Learning. PMLR, 2022. p. 19310-19327. - https://proceedings.mlr.press/v162/scaman22a.html
ROBIN, David, SCAMAN, Kevin, LELARGE, Marc. Convergence beyond the over-parameterized regime using Rayleigh quotients. Poster. NeurIPS, 2022. -

Codes MSC

68T05 Learning and adaptive systems

Dernières questions liées sur MathOverflow

Pour poser une question, votre compte Carmin.tv doit être connecté à mathoverflow

Poser une question sur MathOverflow

Copyright Carmin.tv 2026

Donner son avis