A Fistful of Words: Learning Transferable Visual Models from Bag-of-Words Supervision

Thumbnail Image

Files

2112.13884.pdf (3.53 MB)

Links to Files

https://arxiv.org/abs/2112.13884

Permanent Link

https://doi.org/10.48550/arXiv.2112.13884
 http://hdl.handle.net/11603/26319

Collections

UMBC Computer Science and Electrical Engineering Department
UMBC Faculty Collection

Author/Creator

Author/Creator ORCID

Date

2022-01-06

Type of Work

journal articles
preprints
Text

Department

Program

Citation of Original Publication

Rights

This item is likely protected under Title 17 of the U.S. Copyright Law. Unless on a Creative Commons license, for uses protected by Copyright Law, contact the copyright holder or the author.

Subjects

Abstract

Using natural language as a supervision for training visual recognition models holds great promise. Recent works have shown that if such supervision is used in the form of alignment between images and captions in large training datasets, then the resulting aligned models perform well on zero-shot classification as downstream tasks2 . In this paper, we focus on teasing out what parts of the language supervision are essential for training zero-shot image classification models. Through extensive and careful experiments, we show that: 1) A simple Bag-of-Words (BoW) caption could be used as a replacement for most of the image captions in the dataset. Surprisingly, we observe that this approach improves the zero-shot classification performance when combined with word balancing. 2) Using a BoW pretrained model, we can obtain more training data by generating pseudo-BoW captions on images that do not have a caption. Models trained on images with real and pseudo-BoW captions achieve stronger zero-shot performance. On ImageNet-1k zero-shot evaluation, our best model, that uses only 3M image-caption pairs, performs on-par with a CLIP model trained on 15M image-caption pairs (31.5% vs 31.3%)