Pré-Publication, Document De Travail Année : 2025

Adaptive Class Aware Memory Selection and Contrastive Representation Learning for Robust Online Continual Learning in both Balanced and Imbalanced Data Environments

Résumé

Online Continual Learning (OCL) is a framework where models learn continuously from a stream of data without revisiting previously seen data. This is crucial for many reallife applications, e.g., autonomous driving, healthcare monitoring, and robotics, where data evolves over time. However, current state-of-the-art continuous learning methods struggle with dynamic and unbalanced data, often failing to adapt and leading to severe performance degradation. In this paper, we introduce Memory Selection with Contrastive Learning (MSCL), an advanced approach to Continual Learning (CL) designed to tackle these challenges. MSCL integrates Feature-Distance Based Sample Selection (FDBS) for effective memory adaptation, emphasizing inter-class similarities and intra-class diversity, with a novel contrastive learning loss (SCL) for evolving data representation consolidation. Our extensive evaluations on datasets including CIFAR-100, Mini-ImageNet, PACS, and DomainNet demonstrate that MSCL not only surpasses existing memorybased CL methods on data balanced scenarios, but also excels particularly in imbalanced scenarios, thereby establishing a novel state of the art in both balanced and imbalanced learning contexts. Additionally, we carefully conduct ablation studies to highlight the contribution of each component, i.e., FDBS and SCL, and analyze the impact of the key hyperparameter, i.e., memory size, on the performance of the proposed MSCL method.

Fichier principal
Vignette du fichier
IEEE_neuralsystem_underreview.pdf (558.65 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04929086 , version 1 (04-02-2025)

Licence

Identifiants

  • HAL Id : hal-04929086 , version 1

Citer

Rui Yang, Matthieu Grard, Emmanuel Dellandréa, Liming Chen. Adaptive Class Aware Memory Selection and Contrastive Representation Learning for Robust Online Continual Learning in both Balanced and Imbalanced Data Environments. 2025. ⟨hal-04929086⟩
250 Consultations
332 Téléchargements

Partager

  • More