A masking, linkage and guidance framework for online class incremental learning

Guoqiang Liang, Zhaojie Chen, Shibin Su, Shizhou Zhang, Yanning Zhang

Abstract: Highlights•We design a batch-level logit mask and feature cross fusion module.•An accumulative mean feature distillation module is designed to enhance plasticity.•Based on above modules, we build a masking, linkage and guidance framework for OCIL.•Extensive experiments on popular benchmarks validate the superiority of our method.