Euclidean Geodesic Alignment Prevents Vector Search Collapse on Unseen Classes
A new approach called Euclidean Geodesic Alignment (EGA) has been introduced by researchers as a residual adapter designed for static vision encoders in vector search applications. Current adapters face challenges with out-of-distribution performance, where high-capacity models utilizing global contrastive losses misclassify unseen-class samples into incorrect seen-class clusters, leading to a decline in worst-case Label Precision by more than 40 points compared to the frozen baseline. EGA employs zero initialization, local triplet loss, and hypersphere projection, establishing a self-limiting dynamic that halts gradient production for triplets meeting a minimal margin. Upon convergence, 96.5% of triplets remain gradient-free, preserving unseen-class areas while allowing complete refinement of seen classes. This method is elaborated in arXiv:2605.05674.
Key facts
- Vector search systems using frozen vision encoders face queries from unseen classes at deployment.
- Existing adapter training collapses under distribution shift.
- High-capacity adapters with global contrastive losses drop worst-case Label Precision by over 40 points.
- EGA stands for Euclidean Geodesic Alignment.
- EGA is a residual adapter with three principles: zero initialization, local triplet loss, hypersphere projection.
- The adapter automatically stops updating where local geometry is already correct.
- At convergence, 96.5% of triplets are gradient-free.
- EGA leaves unseen-class regions largely untouched while refining seen classes.
Entities
Institutions
- arXiv