CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning

TMLR Paper6665 Authors

Abstract: Catastrophic forgetting remains a fundamental challenge in continual learning, in which models often forget previous knowledge when fine-tuned on a new task. This issue is especially pronounced in class incremental learning (CIL), which is the most challenging setting in continual learning. Existing methods to address catastrophic forgetting often sacrifice either model interpretability or accuracy. To address this challenge, we introduce Class-Incremental Concept Bottleneck Model (CI-CBM), which leverage effective techniques, including concept regularization and pseudo-concept generation to maintain interpretable decision processes throughout incremental learning phases. Through extensive evaluation on seven datasets, CI-CBM achieves comparable performance to black-box models and outperforms previous interpretable approaches in CIL, with an average 36\% accuracy gain. CI-CBM provides interpretable decisions on individual inputs and understandable global decision rules, as shown in our experiments, thereby demonstrating that human-understandable concepts can be maintained during incremental learning without compromising model performance. Our approach is effective in both pretrained and non-pretrained scenarios; in the latter, the backbone is trained from scratch during the first learning phase.

Submission Type: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

In the revised manuscript (all additions/edits highlighted in blue), we strengthened positioning and evidence in three main ways:

(i) **Expanded related work and comparisons.** We now explicitly discuss other interpretable exemplar-free CIL/CBM methods (e.g., CONCIL and CLG-CBM) and clarify connections to feature-translation approaches such as FeTrIL++.

(ii) **Improved method clarity and reproducibility.** We added implementation and protocol details, including Algorithm 1, and expanded our description of the concept deduplication filter used to avoid redundant concepts during expansion.

(iii) **Added targeted new ablations and analyses.** We include: interpretability without sparsity (Fig. A1); interpretability under reduced concept availability (Figs. A2--A3); an alternative pseudo-sample variant showing concept-space prototype translation is unreliable (Table A8); sensitivity of performance to the distillation weight (Table A10); concept-fidelity metrics validating the distillation regularizer (Table A11); and SigLIP vs. CLIP effects on both accuracy and concept fidelity (Tables A12--A13).

Assigned Action Editor: ~Dmitry_Kangin1

Submission Number: 6665