US 11,874,866 B2
Multiscale quantization for fast similarity search
Xiang Wu, Piscataway, NJ (US); David Simcha, Jersey City, NJ (US); Daniel Holtmann-Rice, New York, NY (US); Sanjiv Kumar, Jericho, NY (US); Ananda Theertha Suresh, New York, NY (US); Ruiqi Guo, Elmhurst, NY (US); and Xinnan Yu, Forest Hills, NY (US)
Assigned to GOOGLE LLC, Mountain View, CA (US)
Filed by Google LLC, Mountain View, CA (US)
Filed on Dec. 14, 2022, as Appl. No. 18/081,376.
Application 18/081,376 is a continuation of application No. 16/638,802, granted, now 11,531,695, previously published as PCT/US2018/032550, filed on May 14, 2018.
Claims priority of provisional application 62/549,094, filed on Aug. 23, 2017.
Prior Publication US 2023/0123941 A1, Apr. 20, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 16/33 (2019.01); G06F 16/31 (2019.01); G06N 20/00 (2019.01)
CPC G06F 16/3347 (2019.01) [G06F 16/313 (2019.01); G06F 16/319 (2019.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method for operating a search engine that is associated with a database that includes a set of items and a rotated residual dataset that includes a set of rotated residuals corresponding to the set of items, the method comprising:
receiving, at a computing system, a search query;
performing, at the computing system, a nearest neighbor search over a quantized encoding of the database based on the search query, wherein each rotated residual of the set of rotated residuals includes a unit-normalized direction component and a scale component, the unit-normalized direction component of each rotated residual being normalized such that a variance of norms of the rotated residual dataset is reduced, and the quantized encoding of the database is based on a quantization of the unit-normalized direction component and a quantization of the scale component of each rotated residual of the set of rotated residuals; and
obtaining, at the computing system, search results based on the nearest neighbor search over the quantized encoding of the database, wherein the search results include an indication of at least a first item of the set of items.