Abstract:Although the concept of Catastrophic Forgetting (CF) is straightforward in Class Incremental Learning (CIL), the causes of CF in models are still vague. In this paper, by introducing the metrics of representation quality, we investigate and analyze systematically, yet quantitatively, the three primary causes for the catastrophic forgetting of model in CIL, namely Intra-phase Forgetting (IpF), Inter-phase Confusion (IpC) and Classifier Deviation (CD). The (i) IpF happens when the learner fails to correctly align the same-phase data as training proceeds, (ii) IpC happens when the learner confuses the current-phase data with the previous, and (iii) CD happens when the old classifier deviates from the current representation space.Â By studying extensively, we have discovered that the current distillation process can effectively address the IpF issue, and NN-based techniques have brought some alleviation to CD. Nevertheless, the IpC dilemma still awaits further exploration. To spark a more in-depth inquiry into this highly relevant question, we propose a simple yet effective framework, \textbf{C}ontrastive \textbf{C}lass \textbf{C}oncentration for \textbf{CIL}~(C4IL). It utilizes the benefit of distillation and contrastive learning, producing a representation distribution that is more cohesive within the same class and clearly different across different classes, in order to reduce representation overlapping throughout the various training phases and address the issue of IpC. Quantitative experiments showcase the effectiveness of our framework: it outperforms the baseline methods iCaRL by 10\% in top-1 accuracy at the final training phase, and the average accuracy is competitive with the up-to-date method RMM. Qualitative results also demonstrate that C4IL significantly lowers the probability of inter-phase confusion and generates a more compact representation distribution which alleviates the IpC problem.

Rotation Augmented Distillation for Exemplar-Free Class Incremental Learning with Detailed Analysis

DiffClass: Diffusion-Based Class Incremental Learning

Multi-granularity knowledge distillation and prototype consistency regularization for class-incremental learning

Class Incremental Learning Via Dynamic Regeneration with Task-Adaptive Distillation

PASS++: A Dual Bias Reduction Framework for Non-Exemplar Class-Incremental Learning

DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free Class-Incremental Learning

Dynamic Residual Classifier for Class Incremental Learning

Context-aware Feature Reconstruction for Class-Incremental Anomaly Detection and Localization

Dual Balanced Class-Incremental Learning With im-Softmax and Angular Rectification

Complementary Learning Subnetworks for Parameter-Efficient Class-Incremental Learning

Hybrid Rotation Self-Supervision and Feature Space Normalization for Class Incremental Learning

G-ACIL: Analytic Learning for Exemplar-Free Generalized Class Incremental Learning

An Analysis of Initial Training Strategies for Exemplar-Free Class-Incremental Learning

On Representation-Level Forgetting in Class Incremental Learning: What's the Bottleneck?

REAL: Representation Enhanced Analytic Learning for Exemplar-free Class-incremental Learning

Gradient Reweighting: Towards Imbalanced Class-Incremental Learning

Non-Exemplar Class-Incremental Learning Via Adaptive Old Class Reconstruction

Rebalancing network with knowledge stability for class incremental learning

Towards Non-Exemplar Semi-Supervised Class-Incremental Learning

Class-Incremental Learning Based on Anomaly Detection