Continual Contrastive Learning: Buffer Cost

This work considers the problem of continual learning (CL) in decentralized scenarios, which has recently attracted attention of the research community for enabling progressive learning in non-stationary environments. Specifically, we focus on the joint Class and Domain Continual Learning (CDCL) setting, where local data might include multiple classes belonging to different domains. Throughout this work, we conduct an ablation study of the state-of-the-art CL methods designed for contrastive learning, highlighting their strengths and weaknesses. Our analysis considers two approaches: Co2L, which leverages contrastive learning, and FNC<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, inspired by neural collapse concepts. Our initial experiments on digits datasets (MNIST and USPS) show that employing contrastive learning (without any replay buffer) is sufficient to learn acceptable representations. However, this is no longer the case when considering more complex datasets (e.g. CIFAR10 or Office-Caltech datasets). In this setting, employing concepts borrowed from neural collapse context becomes useful to enforce learning structured representations.