LACE: Lightweight Attribution-guided Concept Evolution for Continual Learning

Zhiyu Zhu, Jiayu Zhang, Zhibo Jin, Yihong Fei, Xinyi Wang, Huijie Xu, Fang Chen, Jianlong Zhou

Keywords: Interpretability, Continual Learning

Abstract: We address concept proliferation in interpretable continual learning and present LACE, a lightweight framework that couples Concept Bottleneck Models with a learnable concept-alignment layer, \emph{Concept Attribution} (CA) that quantifies per-concept importance under standard attribution axioms, and Concept Verification (CV) that selects pruning budgets via a data-reusable approximation to leave-one-out with an IRLS hat-matrix correction. A prototype-augmentation mechanism stabilizes learning without exemplars. Across coarse- and fine-grained benchmarks, LACE yields compact, reusable concept sets, consistently improves or matches strong baselines, and narrows the gap between average and last-task accuracy, offering an auditable and parameter-efficient route to continual concept evolution. Our code is available at: https://anonymous.4open.science/r/LACE-7FD6/.

Primary Area: interpretability and explainable AI

Submission Number: 2657