EGA adapter induces gradient sparsity so that 96.5% of triplets become inactive at convergence, preserving unseen-class geometry while refining seen classes and improving worst-case Label Precision on OOD benchmarks.
Sigmoid loss for language image pre-training
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
EGA: Adapting Frozen Encoders for Vector Search with Bounded Out-of-Distribution Degradation
EGA adapter induces gradient sparsity so that 96.5% of triplets become inactive at convergence, preserving unseen-class geometry while refining seen classes and improving worst-case Label Precision on OOD benchmarks.