Resolving Task Confusion in Dynamic Expansion Architectures for Class Incremental Learning

The dynamic expansion architecture is becoming popular in class incremental learning, mainly due to its advantages alleviating catastrophic forgetting. However, task confu- sion not well assessed within this framework, e.g., the discrepancy between classes of different tasks learned (i.e., inter-task confusion, ITC), and certain prior- ity still given latest batch old-new con- fusion, ONC). We empirically validate side effects two types confusion. Meanwhile, a novel solution called Task Correlated Incremental Learning (TCIL) pro- posed encourage discriminative fair feature utilization across tasks. TCIL performs multi-level knowledge distil- lation propagate from old new one. It establishes information flow paths at both fea- ture logit levels, enabling learning be aware classes. Besides, attention mechanism classifier re- scoring are applied generate more classification scores. conduct extensive experiments on CIFAR100 Ima- geNet100 datasets. results demonstrate that sistently achieves state-of-the-art accuracy. mitigates ITC ONC, while showing battle with catas- trophic forgetting even no rehearsal memory reserved. Source code: https://github.com/YellowPancake/TCIL.