Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards more holistic interpretability: A lightweight disentangled Concept Bottleneck Model

Gaoxiang Huang, Songning Lai, Yutao Yue

TL;DR

This work addresses input-to-concept mapping bias in Concept Bottleneck Models by proposing a lightweight disentangled CBM (LDCBM) that automatically groups visual features into semantically meaningful components without region annotations. It introduces a filter grouping loss and concept supervision to align grouped visual patterns with human concepts, followed by a concept-to-class predictor; the model is trained with a combined objective $\mathcal{L}_{\text{total}} = \mathcal{L}_y + \lambda_c \mathcal{L}_c + \lambda_g \mathcal{L}_g$, where $\mathcal{L}_g$ enforces intra-group coherence and inter-group separation. Across three datasets (CUB, CelebA, AwA2), LDCBM achieves high concept accuracy (>$90\%$) and competitive or superior class accuracy, with the hybrid CEM+LDCBM showing the strongest gains on CelebA. The results demonstrate improved interpretability and reliability of decisions grounded in visual evidence, offering a practical, scalable path to more responsible interpretable AI.

Abstract

Concept Bottleneck Models (CBMs) enhance interpretability by predicting human-understandable concepts as intermediate representations. However, existing CBMs often suffer from input-to-concept mapping bias and limited controllability, which restricts their practical value, directly damage the responsibility of strategy from concept-based methods. We propose a lightweight Disentangled Concept Bottleneck Model (LDCBM) that automatically groups visual features into semantically meaningful components without region annotation. By introducing a filter grouping loss and joint concept supervision, our method improves the alignment between visual patterns and concepts, enabling more transparent and robust decision-making. Notably, Experiments on three diverse datasets demonstrate that LDCBM achieves higher concept and class accuracy, outperforming previous CBMs in both interpretability and classification performance. By grounding concepts in visual evidence, our method overcomes a fundamental limitation of prior models and enhances the reliability of interpretable AI.

Towards more holistic interpretability: A lightweight disentangled Concept Bottleneck Model

TL;DR

This work addresses input-to-concept mapping bias in Concept Bottleneck Models by proposing a lightweight disentangled CBM (LDCBM) that automatically groups visual features into semantically meaningful components without region annotations. It introduces a filter grouping loss and concept supervision to align grouped visual patterns with human concepts, followed by a concept-to-class predictor; the model is trained with a combined objective , where enforces intra-group coherence and inter-group separation. Across three datasets (CUB, CelebA, AwA2), LDCBM achieves high concept accuracy (>) and competitive or superior class accuracy, with the hybrid CEM+LDCBM showing the strongest gains on CelebA. The results demonstrate improved interpretability and reliability of decisions grounded in visual evidence, offering a practical, scalable path to more responsible interpretable AI.

Abstract

Concept Bottleneck Models (CBMs) enhance interpretability by predicting human-understandable concepts as intermediate representations. However, existing CBMs often suffer from input-to-concept mapping bias and limited controllability, which restricts their practical value, directly damage the responsibility of strategy from concept-based methods. We propose a lightweight Disentangled Concept Bottleneck Model (LDCBM) that automatically groups visual features into semantically meaningful components without region annotation. By introducing a filter grouping loss and joint concept supervision, our method improves the alignment between visual patterns and concepts, enabling more transparent and robust decision-making. Notably, Experiments on three diverse datasets demonstrate that LDCBM achieves higher concept and class accuracy, outperforming previous CBMs in both interpretability and classification performance. By grounding concepts in visual evidence, our method overcomes a fundamental limitation of prior models and enhances the reliability of interpretable AI.
Paper Structure (13 sections, 9 equations, 3 figures, 1 table)

This paper contains 13 sections, 9 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Preliminary
  4. Lightweight Disentangled Concept Bottleneck Models
  5. Learning Disentangled Visual Features
  6. Concept Supervision
  7. Concept-to-Class Prediction
  8. Experiments
  9. Experimental Setup
  10. Result
  11. LDCBM enhance the performance of both concept and class accuracy
  12. LDCBM achieve the better efficiency trade-off Under Intervention
  13. Conclusion and limitation

Figures (3)

  • Figure 1: Case Study: Our proposed LDCBM and vanilla CBM, those t-SNE visualizations of visual pattern cluster relationship learnt in CUB dataset. The different colors in the t-SNE plots and spatial overlays represent distinct feature clusters.
  • Figure 2: Overview of our LDCBM. It includes two stages modules: Alignment of disentangled visual-pattern with concept and Joint training among concept and class. The first forwards to get feature map and cluster for the intra and inter group mask what conduct the filter to concern specific visual pattern. The second do joint train from input to concept, and then mapping the concept to final task classification
  • Figure 3: Intervention result of performing correct and incorrect randomly concept interventions in fine-grained and coarse-grained datasets(CUB and AwA2) respectively by four models. Following kohConceptBottleneckModels2020a, intervention in CUB, we set groups of related concepts together.