Enhancing Concept Localization in CLIP-based Concept Bottleneck Models

TL;DR

The paper tackles the faithfulness of CLIP-based concept bottleneck models (CBMs) by diagnosing concept hallucination, where CLIP activations reflect contextual cues rather than actual image content. It introduces CHILI, a method to disentangle CLIP embeddings into object-related and context-related components while filtering pseudo-register artifacts, enabling localized concept explanations and improved segmentation. Through datasets like ImageNet, MonumAI, and CUB, CHILI demonstrates that the object-focused component $S^{\text{Object}}$ enhances concept detection and segmentation and yields more trustworthy CBMs, with only modest accuracy trade-offs. This work advances interpretable zero-shot CBMs by providing a practical approach to reduce hallucinations and improve faithfulness in explanations, with broader implications for XAI in vision-language models.

Abstract

This paper addresses explainable AI (XAI) through the lens of Concept Bottleneck Models (CBMs) that do not require explicit concept annotations, relying instead on concepts extracted using CLIP in a zero-shot manner. We show that CLIP, which is central in these techniques, is prone to concept hallucination, incorrectly predicting the presence or absence of concepts within an image in scenarios used in numerous CBMs, hence undermining the faithfulness of explanations. To mitigate this issue, we introduce Concept Hallucination Inhibition via Localized Interpretability (CHILI), a technique that disentangles image embeddings and localizes pixels corresponding to target concepts. Furthermore, our approach supports the generation of saliency-based explanations that are more interpretable.

Figures (12)

• Figure 1: Using the CLIP-score between embedings of input images and predefined concepts, labeling-free concept extraction can be performed, allowing prediction on an interpretable latent space (left). However, CLIP tends to hallucinate the presence of concepts, troubling the localisation of CLIP-based CBMs (right).
• Figure 2: Decomposition of the activation map
• Figure 3: Example of explanation produced by the intervention of CHILI in a CBM. On the bottom left, the input image. On the bottom right, the SHAP values. Target label: Orchard Oriole
• Figure 4: Example of explanation produced by the intervention of CHILI in a CBM. On the bottom left, the input image. On the bottom right, the SHAP values. Target label: Brown Creeper
• Figure 5: Example of explanation produced by the intervention of CHILI in a CBM. On the bottom left, the input image. On the bottom right, the SHAP values. Target label: Cape Glossy Starling