US 11,915,118 B2
Method and apparatus for processing computation of zero value in processing of layers in neural network
Saptarsi Das, Bangalore (IN); Sabitha Kusuma, Bangalore (IN); Sehwan Lee, Suwon-si (KR); Ankur Deshwal, Bangalore (IN); and Kiran Kolar Chandrasekharan, Bangalore (IN)
Assigned to Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed by Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed on Feb. 8, 2023, as Appl. No. 18/107,210.
Application 18/107,210 is a continuation of application No. 16/816,861, filed on Mar. 12, 2020, granted, now 11,604,958.
Claims priority of application No. 201941009806 (IN), filed on Mar. 13, 2019; and application No. 10-2020-0010482 (KR), filed on Jan. 29, 2020.
Prior Publication US 2023/0186050 A1, Jun. 15, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/04 (2023.01); G06F 17/16 (2006.01)
CPC G06N 3/04 (2013.01) [G06F 17/16 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method of processing layers in a neural network, the method comprising:
obtaining a plurality of Input Feature Map (IFM) tiles of at least one IFM tensor and a plurality of kernel tiles of at least one kernel tensor from a memory;
performing, by an accelerator, a convolutional operation on the plurality of IFM tiles and the plurality of kernel tiles based on IFM sparsity and kernel sparsity;
generating, by the accelerator, a plurality of partial Output Feature Map (OFM) tiles; and
generating, by the accelerator, a plurality of OFM tiles corresponding to the plurality of IFM tiles using the plurality of partial OFM tiles.