Using OCR Heads to Verbalize Image Semantics

Problem

This work addresses a gap in understanding how visual language models (VLMs) execute optical character recognition (OCR). The authors investigate the mechanisms within VLMs that facilitate the extraction and verbalization of semantic features from images, which is crucial for enhancing the interpretability of these models. The paper is a preprint and has not undergone peer review.

Method

The study examines four VLMs, focusing on the attention heads that are causally necessary for performing OCR tasks. The authors identify general-purpose attention heads that output interpretable semantic features across all image tokens. A novel approach termed the Verbalization Lens Transformation is introduced, which collapses attention weights into a single transformation. This transformation reveals semantic features present in the hidden states across all layers of the model. Furthermore, the method includes a projection to vocabulary space, allowing for the generation of interpretable labels starting from layer 0. An inverse transformation is also employed, enabling the editing of non-word concepts within images, demonstrating the flexibility of the model in manipulating visual content.

Results

The paper provides an output example where the model successfully identifies and verbalizes the concept of "bike" from an image token containing the corresponding visual representation. Additionally, an editing example is presented where a tractor in a naturalistic image is replaced with a revolver, showcasing the model's capability to modify image semantics. However, no quantitative results or baseline comparisons are reported in the text.

Limitations

The authors do not report any limitations in their study. However, the absence of quantitative results and baseline comparisons may hinder the assessment of the method's performance relative to existing approaches.

Why it matters

This research has significant implications for the field of interpretability in VLMs, as it provides insights into the internal mechanisms that enable OCR and semantic feature extraction. By elucidating how attention heads contribute to these processes, the findings could inform future work aimed at enhancing the interpretability and functionality of VLMs in various applications, including image captioning, visual question answering, and other tasks that require a nuanced understanding of visual content.