US 12,033,077 B2
Learning compressible features
Abhinav Shrivastava, Mountain View, CA (US); Saurabh Singh, Mountain View, CA (US); Johannes Ballé, San Francisco, CA (US); Sami Ahmad Abu-El-Haija, East Palo Alto, CA (US); Nicholas Milo Johnston, San Jose, CA (US); and George Dan Toderici, Mountain View, CA (US)
Assigned to Google LLC, Mountain View, CA (US)
Filed by GOOGLE LLC, Mountain View, CA (US)
Filed on Feb. 27, 2023, as Appl. No. 18/175,125.
Application 18/175,125 is a continuation of application No. 16/666,689, filed on Oct. 29, 2019, granted, now 11,610,124.
Application 16/666,689 is a continuation of application No. PCT/US2019/025210, filed on Apr. 1, 2019.
Prior Publication US 2023/0237332 A1, Jul. 27, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06K 9/00 (2022.01); G06F 17/15 (2006.01); G06F 18/24 (2023.01); G06N 3/063 (2023.01); G06N 3/08 (2023.01); G06N 3/082 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 17/15 (2013.01); G06F 18/24 (2023.01); G06N 3/063 (2013.01); G06N 3/082 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
computing, from a dataset, a first set of features that have a first dimension; and
generating a second set of features from the first set of features using a neural network having a set of neural network parameters and a compression method that compresses the first set of features by entropy coding the first set of features using a probability distribution jointly learned with the set of neural network parameters,
wherein the second set of features has a second dimension that is smaller than the first dimension and wherein jointly learning the probability distribution with the set of neural network parameters maximizes a prediction accuracy of the neural network on a prediction task while minimizing a bit rate of the second set of features.