SL-CBM: Enhancing Concept Bottleneck Models with Semantic Locality for Better Interpretability

TL;DR

SL-CBM tackles locality faithfulness in concept bottleneck models by introducing semantic locality through a 1×1 convolution-based concept saliency branch and cross-attention, producing spatially coherent explanations at both concept and class levels. The method remains compatible with pre-trained backbones and uses a multi-term loss to balance accuracy, sparsity, and faithfulness. Extensive experiments on RIVAL-10 and CUB-200-2011 show improved locality faithfulness, explainability, and intervention efficacy without sacrificing predictive performance, with ablations highlighting the importance of entropy regularization and an optional contrastive term. Overall, the work demonstrates that grounding concept reasoning in spatial locality enhances debuggability and paves the way for more trustworthy, interpretable concept-based models.

Abstract

Explainable AI (XAI) is crucial for building transparent and trustworthy machine learning systems, especially in high-stakes domains. Concept Bottleneck Models (CBMs) have emerged as a promising ante-hoc approach that provides interpretable, concept-level explanations by explicitly modeling human-understandable concepts. However, existing CBMs often suffer from poor locality faithfulness, failing to spatially align concepts with meaningful image regions, which limits their interpretability and reliability. In this work, we propose SL-CBM (CBM with Semantic Locality), a novel extension that enforces locality faithfulness by generating spatially coherent saliency maps at both concept and class levels. SL-CBM integrates a 1x1 convolutional layer with a cross-attention mechanism to enhance alignment between concepts, image regions, and final predictions. Unlike prior methods, SL-CBM produces faithful saliency maps inherently tied to the model's internal reasoning, facilitating more effective debugging and intervention. Extensive experiments on image datasets demonstrate that SL-CBM substantially improves locality faithfulness, explanation quality, and intervention efficacy while maintaining competitive classification accuracy. Our ablation studies highlight the importance of contrastive and entropy-based regularization for balancing accuracy, sparsity, and faithfulness. Overall, SL-CBM bridges the gap between concept-based reasoning and spatial explainability, setting a new standard for interpretable and trustworthy concept-based models.

Figures (6)

• Figure 1: Saliency maps of state-of-the-art CBMs and SL-CBM at both concept and class levels. Saliency maps of PCBM yuksekgonul2022post and CSS selvaraj2024improving are generated using GradCAM selvaraju2017grad, while SL-CBM produces its own saliency maps.
• Figure 2: SL-CBM Overview: Given an input image $\mathbf{x}$, a concept set, the fixed concept and image encoders extract concept and image features, and an image feature summary. Projecting the image summary onto concept features yields a similarity vector. A $1 \times 1$ convolution generates concept saliency maps $S_\mathbf{c}$, which, with the similarity vector, are refined via cross-attention into $f(\mathbf{x})$. preserving locality and concept relevance. A classifier then produces logit $g(f(\mathbf{x}))$, and class saliency map $S_l$ is computed by weighting $S_\mathbf{c}$ with the class-specific FCL weight $\mathbf{w}_l$.
• Figure 3: We present the saliency maps on RIVAL-10, showing the saliency maps for the top 5 concepts, where $c^1$ represents the top predicted concept, $c^2$ the second top, and so on, along with the predicted class $l$. The CBM saliency maps are generated using GradCAM selvaraju2017grad, while the SL-CBM saliency maps are produced by our proposed method.
• Figure 4: Ablation study on $\lambda_e$ analyzing its effect on class accuracy, concept accuracy, IoU, Dice, and C-IoU with/without $\mathcal{L}_c$ ($\lambda_c = 1$ or $0$). Experiments are on RIVAL-10 with $\lambda_{ce} = 1$ and $\lambda_{ca} = 10^4$. $\uparrow$ signifies that a higher value is preferable for the metric, while $\downarrow$ indicates that a lower value is better. All values are presented as percentages.
• Figure 5: We present a top 1 concept saliency map on RIVAL-10 of $\lambda_c = 0 \text{ or } 1$.