[Paper Review] Visual Explanations of Image-Text Representations via Multi-Modal Information Bottleneck Attribution

Ying Wang, Tim G. J. Rudner+1 more|arXiv (Cornell University)|Dec 28, 2023

Topic Modeling4 citations

TL;DR

This paper proposes Multi-Modal Information Bottleneck (M2IB) attribution, an information-theoretic method that explains vision-language representations by compressing irrelevant features while preserving relevant visual and textual information. M2IB outperforms gradient-based, perturbation-based, and attention-based methods in attribution accuracy and interpretability, especially in safety-critical domains like healthcare, without requiring ground-truth labels.

ABSTRACT

Vision-language pretrained models have seen remarkable success, but their application to safety-critical settings is limited by their lack of interpretability. To improve the interpretability of vision-language models such as CLIP, we propose a multi-modal information bottleneck (M2IB) approach that learns latent representations that compress irrelevant information while preserving relevant visual and textual features. We demonstrate how M2IB can be applied to attribution analysis of vision-language pretrained models, increasing attribution accuracy and improving the interpretability of such models when applied to safety-critical domains such as healthcare. Crucially, unlike commonly used unimodal attribution methods, M2IB does not require ground truth labels, making it possible to audit representations of vision-language pretrained models when multiple modalities but no ground-truth data is available. Using CLIP as an example, we demonstrate the effectiveness of M2IB attribution and show that it outperforms gradient-based, perturbation-based, and attention-based attribution methods both qualitatively and quantitatively.

Motivation & Objective

• To improve interpretability of vision-language pretrained models (VLPMs) such as CLIP in safety-critical applications like healthcare.
• To develop a multi-modal attribution method that does not rely on ground-truth labels, enabling auditability of representations when labels are unavailable.
• To apply the information bottleneck principle to jointly compress irrelevant features across vision and language modalities while preserving task-relevant information.
• To empirically validate M2IB’s superiority over gradient-based, perturbation-based, and attention-based attribution methods in both qualitative and quantitative evaluations.

Proposed method

• Formulate a multi-modal information bottleneck principle that maximizes the likelihood of one modality’s features given the other modality’s features.
• Use a variational approximation to derive a tractable optimization objective from the information bottleneck principle.
• Optimize with respect to attribution parameters that identify and suppress irrelevant features, generating saliency maps for images and text.
• Insert the information bottleneck after a selected layer (e.g., layer 9) in the model to enable modality-specific attribution.
• Apply the method to CLIP and extend it to other models like ImageBind, demonstrating compatibility with diverse multi-modal architectures.
• Use the ROAR+ score to systematically select optimal hyperparameters and evaluate robustness.

Experimental results

Research questions

• RQ1 Can an information-theoretic attribution method improve interpretability of vision-language models without requiring ground-truth labels?
• RQ2 How does M2IB compare to gradient-based, perturbation-based, and attention-based attribution methods in localizing relevant image and text features?
• RQ3 Does M2IB produce saliency maps that are sensitive to model weight changes, indicating reliable attribution?
• RQ4 Can M2IB be generalized to multi-modal models beyond vision and language, such as those using audio, depth, or thermal data?
• RQ5 How do hyperparameter choices affect the quality and consistency of M2IB attribution maps?

Key findings

• M2IB significantly outperforms gradient-based, perturbation-based, and attention-based attribution methods in both qualitative and quantitative evaluations on image-text attribution tasks.
• M2IB produces more accurate and localized saliency maps, especially in identifying entire relevant objects in images, compared to baseline methods.
• The method passes the sanity check: attribution scores change meaningfully with model weight perturbations, confirming sensitivity to model behavior.
• M2IB generates higher-quality attribution maps for fine-tuned models than for pretrained models, reflecting improved alignment with downstream task features.
• The method is effective on multi-modal models beyond vision and language, as demonstrated on ImageBind with six different modalities, including audio and thermal data.
• Hyperparameter selection significantly affects attribution quality, and the ROAR+ score is effective for identifying optimal configurations.