Conceptlinear probe / 线性分类器探针
Linear Probe
A lightweight interpretability tool that trains a linear classifier on top of model activations to predict some property, such as sentiment or grammaticality, with the weights treated as the direction of that property in activation space.
Timeline (last 90 days)
Oct 14
通过线性探针分析，客户-目标归属信息在模型中间层变得高度可恢复，并且集中在特定的注意力头上。
Unverified50%
Oct 14
线性探针是一种可解释性分析技术，通过在模型某层的激活上训练简单线性分类器，以分类准确率度量该层是否线性编码了某种属性。
Unverified50%