BriefGPT.xyz
Apr, 2024
视觉电路的自动发现
Automatic Discovery of Visual Circuits
HTML
PDF
Achyuta Rajaram, Neil Chowdhury, Antonio Torralba, Jacob Andreas, Sarah Schwettmann
TL;DR
基于视觉概念的神经元激活依赖和功能连接，我们提出了一种新的方法来提取深度视觉模型计算图的子图，从而防御大规模预训练模型的对抗攻击。
Abstract
To date, most discoveries of
network subcomponents
that implement human-interpretable computations in
deep vision models
have involved close study of single units and large amounts of human labor. We explore scal
→