Incrementally Learned Angular Representations for Few-Shot Class-Incremental Learning

The main challenge of FSCIL is the trade-off between underfitting to a new session task and preventing forgetting the knowledge for earlier sessions. In this paper, we reveal that the angular space occupied by the features within the embedded area is relatively narrow. Consequently, after the base session training with the converged feature extractor, adding features of new classes easily overlaps with the previously occupied space of previous classes. Furthermore, in contrast to the base session classes, whose features are relatively well aggregated, the features of new session classes are dispersed over large regions. Thus, we propose the Incrementally Learned Angular Representation (ILAR) learning structure to address these issues. During the base session, ILAR attempts to increase the marginal spacing between the distributions of features for each class. In incremental sessions, modifications of features and classifiers corresponding to the previous session classes are limited to preserve the past knowledge while the features of new classes are readjusted. Furthermore, we generate additional features according to the statistics learned from the base session to reduce the variance caused by data shortage. Experiments proceeded on three popular benchmark datasets, including CIFAR100, miniImageNet, and CUB200. We demonstrate that the proposed method achieves state-of-the-art performances by effectively enhancing the new session classification ability while preserving the knowledge of the past sessions.