Substance or Style: What Does Your Image Embedding Know?

TL;DR

This paper systematically investigates the non-semantic information in image embeddings from various models, revealing differences based on training algorithms and their suitability for style and transformation recognition tasks.

Contribution

It introduces a systematic transformation prediction task to analyze non-semantic content in image embeddings across multiple models, highlighting the impact of training methods.

Findings

Six embeddings encode non-semantic transformation information.

CLIP and ALIGN outperform masking-based models in style transfer recognition.

Model training algorithms influence the types of information captured in embeddings.

Abstract

Probes are small networks that predict properties of underlying data from embeddings, and they provide a targeted, effective way to illuminate the information contained in embeddings. While analysis through the use of probes has become standard in NLP, there has been much less exploration in vision. Image foundation models have primarily been evaluated for semantic content. Better understanding the non-semantic information in popular embeddings (e.g., MAE, SimCLR, or CLIP) will shed new light both on the training algorithms and on the uses for these foundation models. We design a systematic transformation prediction task and measure the visual content of embeddings along many axes, including image style, quality, and a range of natural and artificial transformations. Surprisingly, six embeddings (including SimCLR) encode enough non-semantic information to identify dozens of transformations. We also consider a generalization task, where we group similar transformations and hold out several for testing. We find that image-text models (CLIP and ALIGN) are better at recognizing new examples of style transfer than masking-based models (CAN and MAE). Overall, our results suggest that the choice of pre-training algorithm impacts the types of information in the embedding, and certain models are better than others for non-semantic downstream tasks.