US 11,853,873 B2
Neural network optimization
Daren Croxford, Swaffham Prior (GB); and Jayavarapu Srinivasa Rao, Cambridge (GB)
Assigned to Arm Limited, Cambridge (GB)
Filed by Apical Limited, Cambridge (GB); and Arm Limited, Cambridge (GB)
Filed on Oct. 4, 2018, as Appl. No. 16/152,348.
Prior Publication US 2020/0110995 A1, Apr. 9, 2020
Int. Cl. G06N 3/08 (2023.01); G06N 3/04 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/04 (2013.01)] 25 Claims
OG exemplary drawing
 
1. A method of reducing kernel computations; the method comprising:
ordering a plurality of kernel channels to generate a set of ordered kernel channels based on a potential contribution of each of the plurality of kernel channels, the set of ordered kernel channels having an order of highest potential contribution to lowest potential contribution, and wherein the potential contribution of a given kernel channel is based on an absolute sum of weights associated with the given kernel channel and a given output of the kernel channel when used to process a given input, such that a first kernel channel of the set of ordered kernel channels has the largest potential contribution;
convolving the first kernel channel of the set of ordered kernel channels with input data to produce a convolution output; and
determining whether to convolve subsequent kernel channels of the set of ordered kernel channels,
wherein the step of determining whether to convolve subsequent kernel channels of the set of ordered kernel channels comprises considering the potential contribution of each of a plurality of the subsequent kernel channels of the set of ordered kernel channels, in combination with the convolution output of the first kernel channel.