Beyond Patches: Mining Interpretable Part-Prototypes for Explainable AI

TL;DR

PCMNet tackles the interpretability gap in deep vision by learning adaptive part-prototypes and organizing them into semantically coherent concepts that support human-aligned explanations. It introduces a three-stage pipeline: unsupervised part-prototype discovery, class-specific concept clustering to form centroids, and concept-activation mining that yields a sparse, interpretable decision vector (CAVs) used for classification. The approach, enabled by Marginal Cluster Center Loss and the Concept Mining Module, achieves strong performance on explainability metrics and robustness to occlusion while maintaining competitive accuracy and efficiency relative to prior prototype-based XAI methods. This work advances AI alignment by providing transparent, controllable, and robust explanations suitable for high-stakes applications.

Abstract

As AI systems grow more capable, it becomes increasingly important that their decisions remain understandable and aligned with human expectations. A key challenge is the limited interpretability of deep models. Post-hoc methods like GradCAM offer heatmaps but provide limited conceptual insight, while prototype-based approaches offer example-based explanations but often rely on rigid region selection and lack semantic consistency. To address these limitations, we propose PCMNet, a part-prototypical concept mining network that learns human-comprehensible prototypes from meaningful image regions without additional supervision. By clustering these prototypes into concept groups and extracting concept activation vectors, PCMNet provides structured, concept-level explanations and enhances robustness to occlusion and challenging conditions, which are both critical for building reliable and aligned AI systems. Experiments across multiple image classification benchmarks show that PCMNet outperforms state-of-the-art methods in interpretability, stability, and robustness. This work contributes to AI alignment by enhancing transparency, controllability, and trustworthiness in AI systems. Our code is available at: https://github.com/alehdaghi/PCMNet.

Figures (12)

• Figure 1: Explanation methods. (a) Heatmap-based methods see the feature backbones as black-box models, and try to locate regions activated by the most important features. (b) Patch-based methods decompose the input into image patches and explain the decision using patch-derived components. (c) PCMNet, in contrast, mines both prototypical and non-prototypical concepts, offering a broader and more interpretable set of regions to explain the model decision.
• Figure 2: (a) Overall architecture of PCMNet. In the first stage, part prototypes are learned from spatial features extracted by the backbone. In the second stage, prototypes are clustered within each class to identify frequently occurring concepts. The center of these clusters is computed as concept prototypes, and the distance between part prototypes and concept clusters determines the activated concepts for the model’s decision-making. This approach enables interpretable and concept-driven classification. (b) Concept Mining Module. First, a part centroid list is generated by applying DBSCANDBSCAN clustering within each class and computing the centers of the resulting clusters. These centroids serve as frequently occurring concept prototypes. Next, to activate the most relevant concept for a given part feature, the similarity (inverse distance) to the centroids is measured. These values are then used as CAVs to inform the model’s decision-making.
• Figure 3: visualizations of activated concept on test samples with matching training samples for Stanford Cars and CUB.
• Figure 4: Classification accuracy and faithfulness for a growing level of occlusion.
• Figure 5: t-SNE visualization of concept-level part features extracted by PCMNet. Colors represent different prototypes, while shapes indicate object classes.