US 12,147,901 B2
Training and application method of a multi-layer neural network model, apparatus and storage medium
Hongxing Gao, Beijing (CN); Wei Tao, Beijing (CN); Tsewei Chen, Tokyo (JP); Dongchao Wen, Beijing (CN); and Junjie Liu, Beijing (CN)
Assigned to Canon Kabushiki Kaisha, Tokyo (JP)
Filed by CANON KABUSHIKI KAISHA, Tokyo (JP)
Filed on Dec. 19, 2019, as Appl. No. 16/721,624.
Claims priority of application No. 201811631475.5 (CN), filed on Dec. 29, 2018.
Prior Publication US 2020/0210844 A1, Jul. 2, 2020
Int. Cl. G06N 3/04 (2023.01); G06N 3/084 (2023.01)
CPC G06N 3/084 (2013.01) [G06N 3/04 (2013.01)] 36 Claims
OG exemplary drawing
 
1. A data processing method using a multi-layer neural network model comprising:
inputting data to the multi-layer neural network model;
setting at least one layer of the multi-layer neural network model in which a number of input feature maps is to be expanded and an expansion multiple of expanding the number of input feature maps;
expanding, for the set at least one layer in the network model, a number of input feature maps of the determined at least one layer in accordance with the set expansion multiple by convolving at least one first filter; and
computing the data in each layer of the network model,
wherein data computation in the set at least one layer is convolution computation of the expanded input feature maps and second filters,
wherein the at least one first filter has been updated based on gradient values of the expanded input feature maps in a back propagation in training of the multi-layer neural network model, and
wherein data computation by using the expanded input feature maps comprises:
reading out input feature maps before expansion a plurality of times and computing the data with each of the read out feature maps.