CLIP Interpretability

The foundational technique for visualizing which image pixels most influence a CLIP model's prediction for a specific text query. By computing the gradient of the cosine similarity score between the text and image embeddings with respect to the input image, a heatmap is generated. High-intensity regions indicate pixels that strongly support the text-image pairing, while low-intensity regions are irrelevant. This method is crucial for debugging visual grounding errors and verifying that the model is looking at semantically correct objects rather than spurious background correlations.

A technique for analyzing the shared, high-dimensional space where CLIP projects both images and text. By applying Singular Value Decomposition (SVD) or Principal Component Analysis (PCA) to the embedding matrix, researchers can identify the dominant semantic axes that structure the space. Key applications include:

• Discovering that a single dimension can encode a binary concept like 'day vs. night'
• Measuring the alignment gap between modalities
• Identifying polysemantic neurons that respond to multiple unrelated concepts This decomposition reveals the latent ontology CLIP learned during pre-training.

For CLIP variants with transformer-based fusion, this method examines the attention weights flowing between image patch tokens and text tokens. Attention head specialization is a common finding: some heads consistently attend from object words to corresponding image regions, while others handle spatial relationships or attributes. By ablating specific attention heads and measuring the drop in zero-shot classification accuracy, researchers can build a functional map of which heads are responsible for compositional understanding versus simple object recognition.

A methodology for testing whether a linear classifier can be trained on CLIP's internal representations to predict a high-level concept. For example, a probe trained on the image encoder's hidden states can predict the presence of 'fur' or 'metallic surface' with high accuracy, even if these concepts were never explicitly labeled. This technique is used to:

• Audit for encoded biases in the representation
• Verify that the model has learned compositional features
• Identify spurious features that the model relies on for classification Probing provides a window into the model's emergent ontology.

A specialized interpretability method that exploits CLIP's known vulnerability to text embedded within images. By systematically overlaying words on images and measuring the shift in the model's predicted probability, researchers can quantify the model's reliance on OCR-like reading behavior versus true visual understanding. This technique reveals that CLIP processes rendered text as a dominant signal, often overriding genuine visual features. It is a critical tool for diagnosing multimodal shortcut learning and evaluating the robustness of vision-language models in production.

A quantitative analysis of the geometric separation between image and text embeddings in CLIP's joint space. Despite being trained for alignment, a measurable modality gap persists: image embeddings occupy a distinct, narrow cone of the hypersphere, while text embeddings are more dispersed. This technique measures the gap using the average cosine similarity between random image-text pairs versus in-modality pairs. Understanding this gap is essential for diagnosing retrieval asymmetry—where text-to-image search performs differently than image-to-text search—and for designing better fusion mechanisms.