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

2025-03-07

AI Summary

This paper addresses catastrophic forgetting in exemplar-free class-incremental learning (EFCIL), caused by semantic shift and decision bias during incremental updates. To tackle this, we propose a dual-projection embedding calibration framework coupled with ridge regression-based classifier reconstruction. First, a learnable dual-projection mechanism—comprising row-space projection and nonlinear transformation—is introduced to dynamically estimate and rectify semantic drift of old-class embeddings in the feature space. Second, leveraging calibrated old-class covariance and prototype statistics, we formulate a ridge regression-driven classifier reconstruction paradigm to mitigate decision bias induced by new-task dominance. Our method operates without storing any historical samples and synergistically integrates knowledge distillation with geometric modeling of the embedding space. Evaluated on multiple benchmarks, it consistently outperforms state-of-the-art EFCIL approaches, improving average accuracy across old and new tasks by 3.2–5.8%, thereby achieving superior continual learning trade-offs.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Learning to Search
• Computer Vision: Learning & Optimization for CV

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

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, thereby 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 classifier training as a reconstruction process. This reconstruction exploits previous information encoded in covariance and prototype of each class after calibration with estimated shift, thereby reducing decision bias. Extensive experiments demonstrate that, across various datasets, DPCR effectively balances old and new tasks, outperforming state-of-the-art EFCIL methods.

Problem

• Estimates semantic shift in embedding space for EFCIL.
• Reduces decision bias towards new tasks in EFCIL.
• Balances old and new knowledge in class-incremental learning.

Innovation

• Dual-projection estimates semantic shift effectively
• Ridge regression reconstructs classifier to reduce bias
• Combines learnable transformation with row-space projection