Tuning approximate nearest neighbor search engines for speed-recall tradeoffs via Lagrange multiplier methods
Topics: AI (Deep Learning), Ranking
This Google patent describes a method for optimizing approximate nearest neighbor (ANN) search engines using Lagrange multiplier methods. The technology automatically tunes quantization-based ANN search systems to achieve optimal speed-recall tradeoffs. When given either a desired search cost or recall target, the system uses Lagrangian-based optimization techniques to tune the search parameters, resulting in excellent performance while requiring minimal manual configuration.