A layer-selection rule based on mutual information and linear-probe accuracy identifies where a concept can be probed most accurately in a neural network.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Concept Probing: Where to Find Human-Defined Concepts (Extended Version)
A layer-selection rule based on mutual information and linear-probe accuracy identifies where a concept can be probed most accurately in a neural network.