Partially Shared Concept Bottleneck Models

TL;DR

This work targets interpretability in vision models via Concept Bottleneck Models (CBMs) and identifies three core shortcomings: weak visual grounding, concept redundancy, and lack of principled metrics. It introduces PS-CBM, a three-stage framework that generates multimodal concepts, applies a Partially Shared Concept Strategy to merge and reassign concepts across classes, and trains a CBM with a post-hoc, task-aware evaluation metric called Concept-Efficient Accuracy (CEA). Across 11 diverse datasets, PS-CBM achieves higher accuracy and CEA while using far fewer concepts than prior CBMs, demonstrating improved interpretability without sacrificing performance. The approach combines LLM-driven semantic grounding with exemplar-based visual cues, adaptive concept merging, and a sparse final classifier to deliver compact, semantically coherent explanations for predictions.

Abstract

Concept Bottleneck Models (CBMs) enhance interpretability by introducing a layer of human-understandable concepts between inputs and predictions. While recent methods automate concept generation using Large Language Models (LLMs) and Vision-Language Models (VLMs), they still face three fundamental challenges: poor visual grounding, concept redundancy, and the absence of principled metrics to balance predictive accuracy and concept compactness. We introduce PS-CBM, a Partially Shared CBM framework that addresses these limitations through three core components: (1) a multimodal concept generator that integrates LLM-derived semantics with exemplar-based visual cues; (2) a Partially Shared Concept Strategy that merges concepts based on activation patterns to balance specificity and compactness; and (3) Concept-Efficient Accuracy (CEA), a post-hoc metric that jointly captures both predictive accuracy and concept compactness. Extensive experiments on eleven diverse datasets show that PS-CBM consistently outperforms state-of-the-art CBMs, improving classification accuracy by 1.0%-7.4% and CEA by 2.0%-9.5%, while requiring significantly fewer concepts. These results underscore PS-CBM's effectiveness in achieving both high accuracy and strong interpretability.

Figures (7)

• Figure 1: Illustration of different concept sharing strategies: (a) Independent, where redundant concepts exist across classes; (b) Globally Shared, where predictions are affected by irrelevant concepts; and (c) Partially Shared (Ours), which reduces the number of concepts while avoiding interference from irrelevant ones.
• Figure 2: Overview of the PS-CBM pipeline. Stage 1: Generate multimodal concepts $\mathcal{S}$ by aligning LLM-derived semantics with exemplar images. Stage 2: Apply the Partially Shared Concept Strategy based on activation patterns to construct a concept-labeled dataset $\mathcal{D}'$. Stage 3: Train a transparent sequential predictor on $\mathcal{D}'$ for interpretable image classification.
• Figure 3: Variations of ACC and CEA with the upper limit $K$ on the number of exclusive concepts per class across datasets.
• Figure 4: Ablation study on different concept bottleneck strategies, comparing their impact on ACC and CEA.
• Figure 5: Case study of PS-CBM predictions. Green highlights correct predictions, yellow denotes partially accurate or ambiguous outputs, and red indicates incorrect results.