Contents i List of Tables ii List of Figures iii I. INTRODUCTION 1 II. RELATED WORKS 4 1 Distillation-based methods 4 2 Rehearsal-based Methods 5 3 Data-Free Methods 6 III. Method 7 1 Notation 7 2 Framework of Incremental Learning 8 2.1 Segmentation Loss 9 2.2 Distillation Loss 10 3 Generation of Fake Images with MOSInversion 10 IV. RESULTS 13 1 Dataset 13 2 Implementation details 13 3 Evaluation 14 3.1 Results 14 3.2 The effect of the quality of generated images on model performance 16 3.3 The effect of the learning sequence 20 V. CONCLUSION 22 References 23 i
