Paper Image

Debiasing image classifiers using text

Published on:

30 November 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Juhyeon Park,

Seokhyeon Jeong,

Taesup Moon

Bullets

Key Details

Proposes TLDR method to debias classifiers with text

Text proxies for images via cross-modal transferability

Generated text filtered to remove noise

Competitive with debiasing methods needing balanced images

AI generated summary

Debiasing image classifiers using text

This paper proposes a new method called TLDR to debias image classifiers using only text, without needing additional balanced image datasets. TLDR generates text with large language models, filters the text, then retrains the classifier's last layer using text embeddings as a proxy for images. Results show TLDR matches or exceeds other state-of-the-art debiasing methods that use extra balanced image data.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up