Adaptive Rentention Correction for Continual Learning

Continual learning, also known as lifelong learning or incremental learning,refers to the process by which a model learns from a stream of incoming dataover time. A common problem in continual learning is the classification layer'sbias towards the most recent task. Traditionally, methods have relied onincorporating data from past tasks during training to mitigate this issue.However, the recent shift in continual learning to memory-free environments hasrendered these approaches infeasible. In this study, we propose a solutionfocused on the testing phase. We first introduce a simple Out-of-Task Detectionmethod, OTD, designed to accurately identify samples from past tasks duringtesting. Leveraging OTD, we then propose: (1) an Adaptive Retention mechanismfor dynamically tuning the classifier layer on past task data; (2) an AdaptiveCorrection mechanism for revising predictions when the model classifies datafrom previous tasks into classes from the current task. We name our approachAdaptive Retention Correction (ARC). While designed for memory-freeenvironments, ARC also proves effective in memory-based settings. Extensiveexperiments show that our proposed method can be plugged in to virtually anyexisting continual learning approach without requiring any modifications to itstraining procedure. Specifically, when integrated with state-of-the-artapproaches, ARC achieves an average performance increase of 2.7the CIFAR-100 and Imagenet-R datasets, respectively.