Paper Image

Distilling robustness from vision-language models

Published on:

2 November 2023

Primary Category:

Machine Learning

Paper Authors:

Andy Zhou,

Jindong Wang,

Yu-Xiong Wang,

Haohan Wang

Bullets

Key Details

Proposes distilling robustness from large vision-language models into smaller models

Uses adversarial training and data augmentation for knowledge transfer

Employs a generative model to create diverse adversarial examples

Achieves gains in out-of-distribution robustness across architectures

AI generated summary

Distilling robustness from vision-language models

This paper proposes a framework to improve model robustness by distilling knowledge from large vision-language models into smaller computer vision models, using adversarial training and data augmentation. The key ideas are leveraging the robust representations of foundation models as teachers, and using a generative model to create more diverse adversarial examples for data augmentation during training. Empirical results demonstrate significant gains in out-of-distribution robustness across model architectures.

Answers from this paper

Comments

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

Sign up to comment on this paper

Sign Up