CCSI: Continual Class-Specific Impression for data-free class incremental learning

Sana Ayromlou, Teresa Tsang, Purang Abolmaesumi, Xiaoxiao Li

Abstract: Highlights•Propose a two-step pipeline for data-free class-incremental learning over medical images.•Investigate the effect of normalization layers in both synthesizing and training steps.•Synthesize high-quality and fidelity data as Continual Class-Specific Impression (CCSI) instead of inaccessible original data.•Propose three novel losses to increase the utility of synthesized data in the training step.•Perform qualitative and quantitative evaluation of our pipeline over five various medical imaging classification datasets.