US 12,229,108 B2
Efficient embedding table storage and lookup
Gaurav Menghani, Santa Clara, CA (US)
Assigned to GOOGLE LLC, Mountain View, CA (US)
Filed by Google LLC, Mountain View, CA (US)
Filed on Dec. 20, 2023, as Appl. No. 18/390,524.
Application 18/161,352 is a division of application No. 17/147,844, filed on Jan. 13, 2021, granted, now 11,599,518, issued on Mar. 7, 2023.
Application 18/390,524 is a continuation of application No. 18/161,352, filed on Jan. 30, 2023, granted, now 11,892,998.
Prior Publication US 2024/0211458 A1, Jun. 27, 2024
Int. Cl. G06F 16/00 (2019.01); G06F 16/215 (2019.01); G06F 16/22 (2019.01); G06F 30/27 (2020.01); G06N 3/04 (2023.01); G06N 3/08 (2023.01); G06F 16/174 (2019.01); H03M 7/30 (2006.01)
CPC G06F 16/2282 (2019.01) [G06F 16/215 (2019.01); G06F 16/2255 (2019.01); G06F 16/3347 (2019.01); G06F 30/27 (2020.01); G06N 3/04 (2013.01); G06N 3/08 (2013.01); G06F 16/1744 (2019.01); G06F 2211/007 (2013.01); G06F 2211/1014 (2013.01); H03M 7/30 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A computer-implemented method for generating efficient embedding tables for machine-learning models, the method comprising:
obtaining, by a processor, input data for generating a set of values corresponding to each of a plurality of embedding features;
compressing, by the processor, each of the values in the set for a respective embedding feature of the plurality of embedding features, wherein the compression is configured to enable independent decompression of each value;
forming, by the processor, an embedding table that includes rows, each row comprising an index and a compressed value corresponding to the index, wherein the embedding table is configured to allow decompression of the compressed value associated with the index independent of other rows in the embedding table; and
providing, by the processor, the embedding table for utilization in a machine-learning model.