Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, Yihong Gong; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 12183-12192

Abstract

  • FSCIL: It requires CNN models to incrementally learn new classes from very few labelled samples, without forgetting the previously learned ones.
  • TOPIC: Topology-preserving knowledge incrementer framework, which is proposed to mitigates the forgetting of the old classes by stabilizing NG’s topology and improves the representation learning for few-shot new classes by growing and adapting NG to new training samples.

Few-Shot Class-Incremental Learning

  • Task Design
    • a stream of training sets $D^{(1)}, D^{(2)}, \dots$, and for each training set $D^{(t)}$, $L^{(t)}$ is the set of classes of $D^{(t)}$.