A concept-based alignment analysis using manifold concepts and a partitioned Rand index reveals that increased supervision reduces semantic structure in late-layer ViT representations.
Network dissection: Quantifying inter- pretability of deep visual representations
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers
A concept-based alignment analysis using manifold concepts and a partitioned Rand index reveals that increased supervision reduces semantic structure in late-layer ViT representations.