Coarse-To-Fine Incremental Few-Shot Learning

AbstractDifferent from fine-tuning models pre-trained on a large-scale dataset of preset classes, class-incremental learning (CIL) aims to recognize novel classes over time without forgetting classes. However, given model will be challenged by test images with finer-grained e.g., basenji is at most recognized as dog. Such form new training set (i.e., support set) so that the incremental hoped query) next time. This paper formulates such hybrid natural problem coarse-to-fine few-shot (C2FS) recognition CIL named C2FSCIL, and proposes simple, effective, theoretically-sound strategy Knowe: learn, freeze, normalize classifier’s weights fine labels, once an embedding space contrastively coarse labels. Besides, stability-plasticity balance, overall performance metrics are proposed. In hat sense, CIFAR-100, BREEDS, tieredImageNet, Knowe outperforms all recent relevant or FSCIL methods.KeywordsClass-incremental learningCoarse-to-fineFew shots