MedSAE neurons extracted from MedCLIP achieve higher monosemanticity and interpretability than raw features on CheXpert, measured via correlation, entropy, and automated naming with MedGemma.
1: The overall proposed pipeline
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders
MedSAE neurons extracted from MedCLIP achieve higher monosemanticity and interpretability than raw features on CheXpert, measured via correlation, entropy, and automated naming with MedGemma.