Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental Learning

TL;DR

This work tackles exemplar-free class-incremental learning (EFCIL), where models forget previously learned knowledge due to semantic shift and decision bias after learning new tasks. It proposes DPCR, a framework that first estimates and compensates semantic shift with a dual-projection mechanism (task-wise semantic shift projection and category information projection) and then reconstructs a balanced classifier using ridge regression (RRCR) that leverages calibrated covariance and class prototypes without replaying data. By integrating shift-calibrated information and a BP-free classifier reconstruction, DPCR achieves a favorable stability-plasticity trade-off and outperforms state-of-the-art EFCIL methods across CIFAR-100, Tiny-ImageNet, and ImageNet-100, among others. The approach provides practical benefits for real-world continual learning applications where storing exemplars is impractical and demonstrates strong scalability to large-scale tasks.

Abstract

Exemplar-Free Class-Incremental Learning (EFCIL) aims to sequentially learn from distinct categories without retaining exemplars but easily suffers from catastrophic forgetting of learned knowledge. While existing EFCIL methods leverage knowledge distillation to alleviate forgetting, they still face two critical challenges: semantic shift and decision bias. Specifically, the embeddings of old tasks shift in the embedding space after learning new tasks, and the classifier becomes biased towards new tasks due to training solely with new data, hindering the balance between old and new knowledge. To address these issues, we propose the Dual-Projection Shift Estimation and Classifier Reconstruction (DPCR) approach for EFCIL. DPCR effectively estimates semantic shift through a dual-projection, which combines a learnable transformation with a row-space projection to capture both task-wise and category-wise shifts. Furthermore, to mitigate decision bias, DPCR employs ridge regression to reformulate a classifier reconstruction process. This reconstruction exploits previous in covariance and prototype of each class after calibration with estimated shift, thereby reducing decision bias. Extensive experiments demonstrate that, on various datasets, DPCR effectively balances old and new tasks, outperforming state-of-the-art EFCIL methods. Our codes are available at https://github.com/RHe502/ICML25-DPCR.

Figures (7)

• Figure 1: After learning new tasks, (a) the embeddings of old tasks undergo semantic shift in embedding space, (b) and the embeddings are more likely to be classified as new classes.
• Figure 2: An overview of our proposed DPCR. (a) At task t, the backbone is first trained with new data to learn new representation. (b) After the representation learning, shift estimation is conducted with dual-projection (DP), which is consisted of TSSP and CIP. (c) With the DP, the RRCR reconstructs the classifier based on calibrated covariance and prototypes.
• Figure 3: Evolution curves of task-wise accuracy.
• Figure 4: Stability-plasticity analysis with and without CIP.
• Figure 5: Balance Effect of Classification Reconstruction on Stability and Plasticity.