Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIP

TL;DR

This paper presents a framework to interpret and decompose the internal components of various vision transformers (ViTs) using text, enabling better understanding and application of these models beyond CLIP.

Contribution

The authors introduce a general method to decompose ViT representations and interpret component roles via text, applicable to multiple ViT variants beyond CLIP.

Findings

Identified roles of ViT components in image feature representation

Enabled text-based interpretation and visualization of ViT components

Improved image retrieval and understanding through component analysis

Abstract

Recent work has explored how individual components of the CLIP-ViT model contribute to the final representation by leveraging the shared image-text representation space of CLIP. These components, such as attention heads and MLPs, have been shown to capture distinct image features like shape, color or texture. However, understanding the role of these components in arbitrary vision transformers (ViTs) is challenging. To this end, we introduce a general framework which can identify the roles of various components in ViTs beyond CLIP. Specifically, we (a) automate the decomposition of the final representation into contributions from different model components, and (b) linearly map these contributions to CLIP space to interpret them via text. Additionally, we introduce a novel scoring function to rank components by their importance with respect to specific features. Applying our framework to various ViT variants (e.g. DeiT, DINO, DINOv2, Swin, MaxViT), we gain insights into the roles of different components concerning particular image features. These insights facilitate applications such as image retrieval using text descriptions or reference images, visualizing token importance heatmaps, and mitigating spurious correlations. We release our code to reproduce the experiments at https://github.com/SriramB-98/vit-decompose