Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty-Aware Concept Bottleneck Models with Enhanced Interpretability

Haifei Zhang, Patrick Barry, Eduardo Brandao

TL;DR

The paper tackles the interpretability-uncertainty trade-off in Concept Bottleneck Models (CBMs) by introducing a Class-Level Prototype Classifier (CLPC) that learns one binary prototype per class in the concept space and classifies via proximity to predicted concepts. Uncertainty is quantified through a conformal prediction framework that yields set-valued predictions and the option to abstain, while interpretability is enhanced by rule-like prototypes and per-concept contribution analyses. The approach includes a two-stage training procedure (concept predictor followed by prototype learning with sparsity and binarization regularizers) and a targeted concept-intervention mechanism that provides counterfactual explanations for misclassifications. Empirical results on CUB-200-2011, Derm7pt, and RIVAL10 show competitive label accuracy with improved uncertainty calibration, robustness to concept-noise, and effective, gain-driven concept interventions. This work advances CBMs by delivering reliable, interpretable, and uncertainty-aware predictions suitable for high-stakes applications.

Abstract

In the context of image classification, Concept Bottleneck Models (CBMs) first embed images into a set of human-understandable concepts, followed by an intrinsically interpretable classifier that predicts labels based on these intermediate representations. While CBMs offer a semantically meaningful and interpretable classification pipeline, they often sacrifice predictive performance compared to end-to-end convolutional neural networks. Moreover, the propagation of uncertainty from concept predictions to final label decisions remains underexplored. In this paper, we propose a novel uncertainty-aware and interpretable classifier for the second stage of CBMs. Our method learns a set of binary class-level concept prototypes and uses the distances between predicted concept vectors and each class prototype as both a classification score and a measure of uncertainty. These prototypes also serve as interpretable classification rules, indicating which concepts should be present in an image to justify a specific class prediction. The proposed framework enhances both interpretability and robustness by enabling conformal prediction for uncertain or outlier inputs based on their deviation from the learned binary class-level concept prototypes.

Uncertainty-Aware Concept Bottleneck Models with Enhanced Interpretability

TL;DR

The paper tackles the interpretability-uncertainty trade-off in Concept Bottleneck Models (CBMs) by introducing a Class-Level Prototype Classifier (CLPC) that learns one binary prototype per class in the concept space and classifies via proximity to predicted concepts. Uncertainty is quantified through a conformal prediction framework that yields set-valued predictions and the option to abstain, while interpretability is enhanced by rule-like prototypes and per-concept contribution analyses. The approach includes a two-stage training procedure (concept predictor followed by prototype learning with sparsity and binarization regularizers) and a targeted concept-intervention mechanism that provides counterfactual explanations for misclassifications. Empirical results on CUB-200-2011, Derm7pt, and RIVAL10 show competitive label accuracy with improved uncertainty calibration, robustness to concept-noise, and effective, gain-driven concept interventions. This work advances CBMs by delivering reliable, interpretable, and uncertainty-aware predictions suitable for high-stakes applications.

Abstract

In the context of image classification, Concept Bottleneck Models (CBMs) first embed images into a set of human-understandable concepts, followed by an intrinsically interpretable classifier that predicts labels based on these intermediate representations. While CBMs offer a semantically meaningful and interpretable classification pipeline, they often sacrifice predictive performance compared to end-to-end convolutional neural networks. Moreover, the propagation of uncertainty from concept predictions to final label decisions remains underexplored. In this paper, we propose a novel uncertainty-aware and interpretable classifier for the second stage of CBMs. Our method learns a set of binary class-level concept prototypes and uses the distances between predicted concept vectors and each class prototype as both a classification score and a measure of uncertainty. These prototypes also serve as interpretable classification rules, indicating which concepts should be present in an image to justify a specific class prediction. The proposed framework enhances both interpretability and robustness by enabling conformal prediction for uncertain or outlier inputs based on their deviation from the learned binary class-level concept prototypes.

Paper Structure

This paper contains 16 sections, 15 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Concept bottleneck models
  3. Class-level prototype classifier
  4. Model structure and inference
  5. Class-level prototype learning
  6. Uncertainty-aware prediction
  7. Interpretability
  8. Concept intervention
  9. Experiments
  10. Label prediction
  11. Robustness to noise in concept prediction
  12. Concept intervention efficiency
  13. Conclusion
  14. Acknowledgments.
  15. Disclosure of Interests.
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Model structure of our proposed class-level prototype network.
  • Figure 2: Visualization of distance decomposition between predicted concept scores and the closest class prototypes
  • Figure 3: Top-1 accuracy with respect to levels of concept noise in test set.
  • Figure 4: Number of concept interventions required to correct misclassified test instances.