A suite of orientation-probing metrics shows that vision transformers encode orientation in a depth profile set by the training objective, and the measured peak predicts the best layer to start fine-tuning.
8), with the decoder rows constrained to unit norm so the sparsity penalty cannot be evaded by rescaling
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
IRIS: A Visual Cortex-Inspired Framework for Analyzing Orientation Selectivity in Vision Transformers
A suite of orientation-probing metrics shows that vision transformers encode orientation in a depth profile set by the training objective, and the measured peak predicts the best layer to start fine-tuning.