Paper Image

Understanding vision models via text

Published on:

16 October 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Haozhe Chen,

Junfeng Yang,

Carl Vondrick,

Chengzhi Mao


Key Details

Proposes interpreting vision model latent tokens with text descriptions

Maps tokens to final layer through local operations, no extra data/training needed

Shows this reveals model's hierarchical reasoning process

Enables controlling model via replacing tokens, e.g. fixing attacks

AI generated summary

Understanding vision models via text

This paper proposes a method to interpret the internal representations in vision transformer models using natural language descriptions. It maps the latent token embeddings to final layer text embeddings through local operations, then retrieves closest text descriptions without extra data or training. This enables understanding models' reasoning and controlling behaviors by replacing tokens, like fixing attacks, reducing bias, and directing reasoning chains.

Answers from this paper


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

Sign up to comment on this paper

Sign Up