You look from old classes: Towards accurate few shot class-incremental learning

Few-shot class incremental learning (FSCIL) is a common but difficult task that faces two challenges: catastrophic forgetting of old classes and insufficient learning of new classes with limited samples. Recent wisdom focuses on preventing catastrophic forgetting yet overlooks the limited samples issue, resulting in poor new class performance. In this paper, we argue that old class samples contain rich knowledge, which can be exploited to supplement the learning of new classes. To this end, we propose to Look from Old Classes (YLOC) for FSCIL, enhancing both the base and incremental sessions. In the base session, we develop a prototype centered loss (PCL) to obtain a compact distribution of old classes. During incremental sessions, we devise a prototype augmentation learning (PAL) method to aid the learning of new classes by exploiting old classes. Extensive experiments on three FSCIL benchmark datasets demonstrate the superiority of our method.