Pseudo-Labeling for Class Incremental Learning - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Pseudo-Labeling for Class Incremental Learning

Résumé

Class Incremental Learning (CIL) consists in training a model iteratively with limited amount of data from few classes that will never be seen again, resulting in catastrophic forgetting and lack of diversity. In this paper, we address these phenomena by assuming that, during incremental learning, additional unlabeled data are continually available, and propose a Pseudo-Labeling approach for class incremental learning (PLCiL) that makes use of a new adapted loss. We demonstrate that our method achieves better performance than supervised or other semi-supervised methods on standard class incremental benchmarks (CIFAR-100 and ImageNet-100) even when a self-supervised pre-training step using a large set of data is used as initialization. We also illustrate the advantages of our method in a more complex context with fewer labels.
Fichier principal
Vignette du fichier
DTIS21234.pdf (441.4 Ko) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte

Dates et versions

hal-03464466 , version 1 (03-12-2021)

Identifiants

  • HAL Id : hal-03464466 , version 1

Citer

Alexis Lechat, Stéphane Herbin, Frédéric Jurie. Pseudo-Labeling for Class Incremental Learning. BMVC 2021 : The British Machine Vision Conference, Nov 2021, virtuel, United Kingdom. ⟨hal-03464466⟩
105 Consultations
240 Téléchargements

Partager

Gmail Facebook X LinkedIn More