@

초록

Interpreting the decisions of deep neural networks remains a significant challenge, as conventional class activation maps (CAMs) primarily highlight class-level cues rather than the underlying concepts that constitute them. In this work, we propose an enhanced framework that extends CAPE by introducing a class-wise concept probability branch. Leveraging Slot Attention, our method discovers a fixed set of interpretable concepts, and a shared multilayer perceptron (MLP) maps these concept embeddings to class scores, resulting in spatial concept maps that are directly human-interpretable. We integrate concept-based evidence with CAPE-based evidence via a learnable gating mechanism, and further modulate the fused explanation by the classifier posterior, ensuring that the final explanation is both spatially precise and consistent with the model's belief. The training objective combines standard classification loss on the original branch, knowledge distillation from the original to the CAPE branch, and optional regularizers that promote intra-concept consistency and inter-concept distinctiveness. Additionally, a completeness objective is employed to encourage the learned concepts to comprehensively account for the model's predictions. Our approach produces concept-level explanations that decompose class evidence into semantically meaningful parts, while preserving the probabilistic structure and interpretability advantages of CAPE.

키워드