US 11,704,571 B2
Learned threshold pruning for deep neural networks
Kambiz Azarian Yazdi, San Diego, CA (US); Tijmen Pieter Frederik Blankevoort, Amsterdam (NL); Jin Won Lee, San Diego, CA (US); and Yash Sanjay Bhalgat, San Diego, CA (US)
Assigned to QUALCOMM Incorporated, San Diego, CA (US)
Filed by QUALCOMM Incorporated, San Diego, CA (US)
Filed on Oct. 9, 2020, as Appl. No. 17/67,233.
Claims priority of provisional application 62/914,233, filed on Oct. 11, 2019.
Prior Publication US 2021/0110268 A1, Apr. 15, 2021
Int. Cl. G06N 3/08 (2023.01); G06N 3/082 (2023.01); G06N 3/04 (2023.01)
CPC G06N 3/082 (2013.01) [G06N 3/04 (2013.01)] 28 Claims
OG exemplary drawing
 
1. A processor-implemented method, comprising:
determining a pruning threshold for pruning a plurality of pre-trained weights of a first artificial neural network (ANN) model based on a function of a classification loss and a regularization loss, the regularization loss comprising a count of unpruned weights;
pruning a first set of pre-trained weights, of the plurality of pre-trained weights of the first ANN model, with a first value that is less than the pruning threshold;
adjusting a second set of pre-trained weights of the plurality of pre-trained weights of the first ANN model in response to a second value of each pre-trained weight in the second set of pre-trained weights being greater than the pruning threshold; and
generating a second ANN model based on the adjusted second set of pretrained weights.