LRP-CLIP: A Zero Shot Approach for the Explanation of the Cognitive Functions of Vision Models

Malte Singerhoff, Viktor Matkovic, Torben Weis

Keywords: LRP, CLIP, VIT, TRANSFORMER, XAI, Cognitive, Zer Shot

TL;DR: LRP-CLIP is a zero-shot framework that combines LRP heatmaps with CLIP to complements models and human interpretation.

Abstract:

Cognitive explanation of deep neural networks is enabled by the use of Layer-wise Relevance Propagation (LRP) heatmaps. Heatmaps indicate where a region of high significance is, but current approaches rely on further computational training overhead to assign semantic meaning. We propose LRP-CLIP, a zero-shot framework that grounds these heatmaps in natural language using pretrained CLIP (Contrastive Language-Image Pre-Training) and thereby allowing cognitive explanations of deep learning models without further training.
Our method extracts relevance-based image crops and matches them against domain-specific textual attributes, producing human-readable explanations without any additional training or supervision. Applied to bird classification with a Vision Transformer (ViT), we find that the model relies heavily on head features, while occasionally exploiting spurious background cues. This dual pattern reflects both systematic heuristics and potential biases in the model’s cognitive strategy. Quantitative evaluation shows that our pipeline reliably localizes annotated parts and assigns semantically valid labels, clearly outperforming random baselines. These results demonstrate an extensible zero-shot approach to model cognition.

Submission Number: 18