Abstract
In this paper, we study the effect of a novel regularization scheme on contrastive language-image pre-trained (CLIP) models. Our approach is based on the observation that, in many domains, text tokens should only describe a small number of image regions and, likewise, each image region should correspond to only a few text tokens. In CLIP-style models, this implies that text-token embeddings should have high similarity to only a small number of image-patch embeddings for a given image-text pair. We formalize this observation using a novel regularization scheme that penalizes the entropy of the text-token to image-patch similarity scores. We qualitatively and quantitatively demonstrate that the proposed regularization scheme shrinks the text-token and image-patch similarity scores towards zero, thus achieving the desired effect. We demonstrate the promise of our approach in an important medical context where this underlying hypothesis naturally arises. Using our proposed approach, we achieve state of the art (SOTA) zero-shot performance on all tasks from the CheXpert chest x-ray dataset, outperforming an unregularized version of the model and several recently published self-supervised models.
Abstract (translated)
URL
https://arxiv.org/abs/2212.06710
