Thinking Like Humans: Explainable Class-Incremental Learning via Distribution Matching
Abstract
Class Incremental Learning (CIL) aims to enable machines to continuously acquire new knowledge without forgetting previously learned concepts. However, existing methods primarily rely on discriminative parameter expansion, making them highly prone to catastrophic forgetting due to parameter overwriting. Furthermore, their opaque black-box nature limits their reliability in high-stakes real-world applications. In contrast, humans learn by matching salient visual details with semantic structures, naturally forming traceable reasoning chains. Inspired by this cognitive process, we propose a novel Explainable Class Incremental Learning(ECIL) paradigm which shifts from opaque parametric memory to explicit, human-like reasoning. We simulate human visual attention through soft-label-guided augmentation to extract key discriminative image regions. Instead of expanding the classification head, we utilize structured semantic attributes as explicit knowledge anchors, fundamentally preventing the forgetting of prior knowledge. To bridge the visual and semantic modalities, we design an entropy-regularized optimal distribution matching (DM) strategy. This mechanism not only inherently suppresses feature interference across tasks but also explicitly quantifies the contribution of specific image regions to semantic attributes, thereby generating transparent reasoning chains for each decision. Extensive experiments demonstrate that our method achieves an excellent balance between high accuracy, low forgetting rates, and intuitive interpretability, paving the way for safe and reliable continuous learning.