PIB regularizes prompt-tuned frozen ViTs with layer-wise compression and sufficiency losses, improving VTAB-1k transfer to 77.33% with about 0.35-0.51% trainable parameters.
Title resolution pending
1 Pith paper cite this work, alongside 3 external citations. Polarity classification is still indexing.
1
Pith paper citing it
3
external citations · OpenAlex
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Rethinking Layer-Wise Information Allocation for Vision Foundation Model Adaptation
PIB regularizes prompt-tuned frozen ViTs with layer-wise compression and sufficiency losses, improving VTAB-1k transfer to 77.33% with about 0.35-0.51% trainable parameters.