SOAR: Improved Indexing for Approximate Nearest Neighbor Search
Topics: AI (Deep Learning), Document Classification, Indexing, Ranking
This Google research paper presents SOAR (Spilling with Orthogonality-Amplified Residuals), a new indexing method for approximate nearest neighbor (ANN) search that builds on redundant “spilled” data representations but optimizes them jointly using an orthogonality-amplified residual loss. By encouraging partitioning residuals to be orthogonal, SOAR mitigates failure-correlation across spilled assignments, yielding significantly higher recall and throughput on large-scale ANN benchmarks—all with minimal increases in indexing time and memory usage.