Lune

CVPR Findings2026

Class-Aware Drift Compensation for Non-Uniform Semantic Shift in Continual Learning

Fankang Xu, Lu Jin, Yanpeng Sun, Shiyu Xuan, Zechao Li

Abstract

Continual learning aims to develop models that can incrementally acquire new knowledge without sacrificing previously acquired information. When the storage of past data is completely prohibited, catastrophic forgetting is exacerbated, stimulating research in exemplar-free continual learning, where class prototypes are typically retained and periodically updated via drift compensation. However, these drift compensation methods generally assume that semantic shift between tasks occurs uniformly (i.e., uniformly mild or uniformly abrupt), whereas in real-world data streams, such shift is often unpredictable. To address this issue, we propose a Class-Aware Drift Compensation (CADC) method to handle feature drift under non-uniform semantic shift. The method dynamically models the drift degree of old classes under the challenging exemplar-free scenario, enabling class-specific adaptive compensation. Furthermore, we demonstrate that the class drift modeled by CADC can effectively guide knowledge distillation by imposing class-specific constraints on logits based on drift intensity, which motivates the design of drift-guided distillation to better balance model stability and plasticity. Extensive experiments on six datasets demonstrate the superiority of the proposed method over existing exemplar-free approaches and the advantage of CADC over state-of-theart drift compensation techniques. Code is available at https://github.com/AldrinLake/CADC.git. Bias induced by excessive or insufficient drift compensation Original prototype Compensated prototype Oracle prototype Traditional drift compensation: uniform across classes Real semantic shifts: diverse across classes 9.03

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

  • FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning cites Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning
  • Prototype Reminiscence and Augmented Asymmetric Knowledge Aggregation for Non-Exemplar Class-Incremental Learning cites Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning
  • Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning cites Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning
Builds onCited byLeft to right by year; the more cited, the higher.

Builds on29

  1. DyTox: Transformers for Continual Learning with DYnamic TOken eXpansionArthur Douillard, Alexandre Ramé, Guillaume Couairon, Matthieu CordCVPR 2022 · 315 citations
  2. Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental LearningKai Zhu, Wei Zhai, Yang Cao, Jiebo Luo et al.CVPR 2022 · 155 citations
  3. FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual LearningDipam Goswami, Yuyang Liu, Bartlomiej Twardowski, Joost van de WeijerNeurIPS 2023 · 136 citations
  4. Prototype Reminiscence and Augmented Asymmetric Knowledge Aggregation for Non-Exemplar Class-Incremental LearningWuxuan Shi, Mang YeICCV 2023 · 49 citations
  5. Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental LearningSimone Magistri, Tomaso Trinci, Albin Soutif-Cormerais, Joost van de Weijer et al.ICLR 2024 · 42 citations

Related papers

  1. Dual-Estimator: Decoupling Global and Local Semantic Shift for Drift Compensation in Class-Incremental LearningFankang Xu, Lu Jin, Yanpeng Sun, Shiyu Xuan et al.CVPR 2026
  2. Resurrecting Old Classes with New Data for Exemplar-Free Continual LearningDipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath et al.CVPR 2024
  3. Bring Evanescent Representations to Life in Lifelong Class Incremental LearningMarco Toldo, Mete OzayCVPR 2022 · 34 citations
  4. Prototype-Sample Relation Distillation: Towards Replay-Free Continual LearningNader Asadi, MohammadReza Davari, Sudhir P. Mudur, Rahaf Aljundi et al.ICML 2023 · 61 citations
  5. TARGET: Federated Class-Continual Learning via Exemplar-Free DistillationJie Zhang, Chen Chen, Weiming Zhuang, Lingjuan LyuICCV 2023 · 109 citations
Dusk over the sea between two cliffs drawn in fine vertical lines