Determine the Necessary Number of Auxiliary Query Vectors

Determine how many auxiliary query vectors are necessary during construction of cross-modal out-of-distribution approximate nearest-neighbor search indexes to achieve sufficient retrieval performance without incurring unnecessary construction overhead.

Background

Existing out-of-distribution approximate nearest-neighbor search methods use auxiliary query vectors to guide graph construction toward the query distribution. The construction cost is therefore sensitive to the number of auxiliary queries ingested. The paper observes that reducing the auxiliary-query set can preserve comparable index performance, while additional queries may mainly produce redundant updates to already well-connected nodes.

The unresolved issue is to identify how many auxiliary queries are actually needed before the graph has acquired sufficient query-distribution information. This motivates CORAL’s incremental construction procedure and its convergence-based early-termination mechanism, but the general number of necessary auxiliary queries remains unspecified.