US 12,248,866 B2
Method and apparatus for reducing computational complexity of convolutional neural networks
Mostafa El-Khamy, San Diego, CA (US); Yoo Jin Choi, San Diego, CA (US); and Jungwon Lee, San Diego, CA (US)
Assigned to Samsung Electronics Co., Ltd, (KR)
Filed by Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed on Sep. 13, 2021, as Appl. No. 17/473,813.
Application 17/473,813 is a continuation of application No. 15/634,537, filed on Jun. 27, 2017, granted, now 11,164,071.
Claims priority of provisional application 62/486,626, filed on Apr. 18, 2017.
Prior Publication US 2021/0406647 A1, Dec. 30, 2021
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/06 (2006.01); G06N 3/04 (2023.01); G06N 3/0464 (2023.01); G06N 3/08 (2023.01); G06F 5/01 (2006.01); G06F 17/16 (2006.01)
CPC G06N 3/06 (2013.01) [G06N 3/04 (2013.01); G06N 3/0464 (2023.01); G06N 3/08 (2013.01); G06F 5/01 (2013.01); G06F 17/16 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A convolutional neural network (CNN) system implemented by one or more computers for generating a classification for an input image received by the CNN system, the CNN system comprising:
circuitry running on clock cycles, the circuitry configured to compute a product of two received values; and
at least one non-transitory computer-readable medium that stores instructions for the circuitry to:
derive a feature map based on at least the input image;
puncture at least one selection among the feature map and a kernel by setting a value of an element at an index of the at least one selection to zero, and cyclic-shifting a puncture pattern to achieve a 1/d reduction in number of clock cycles, where d is an integer and puncture interval value >1;
convolve the feature map with the kernel to generate a first convolved output;
store the first convolved output in a register; and
generate the classification for the input image based on at least the first convolved output.