US 11,907,844 B2
Processing method and accelerating device
Zidong Du, Pudong New Area (CN); Xuda Zhou, Pudong New Area (CN); Shaoli Liu, Pudong New Area (CN); and Tianshi Chen, Pudong New Area (CN)
Assigned to SHANGHAI CAMBRICON INFORMATION TECHNOLOGY CO., LTD, Shanghai (CN)
Filed by Shanghai Cambricon Information Technology Co., Ltd., Pudong New Area (CN)
Filed on Nov. 28, 2019, as Appl. No. 16/699,032.
Application 16/699,032 is a continuation of application No. 16/699,027, filed on Nov. 28, 2019.
Application 16/699,027 is a continuation in part of application No. PCT/CN2018/088033, filed on May 23, 2018.
Prior Publication US 2020/0104693 A1, Apr. 2, 2020
Int. Cl. G06N 3/082 (2023.01); G06F 1/3296 (2019.01); G06F 9/38 (2018.01); G06F 13/16 (2006.01); G06N 3/04 (2023.01); G06N 3/084 (2023.01); G06F 16/28 (2019.01); G06N 3/063 (2023.01); G06F 12/0875 (2016.01); G06N 3/044 (2023.01); G06N 3/048 (2023.01)
CPC G06N 3/082 (2013.01) [G06F 1/3296 (2013.01); G06F 9/3877 (2013.01); G06F 12/0875 (2013.01); G06F 13/16 (2013.01); G06F 16/285 (2019.01); G06N 3/04 (2013.01); G06N 3/044 (2023.01); G06N 3/048 (2023.01); G06N 3/063 (2013.01); G06N 3/084 (2013.01); G06F 2212/452 (2013.01); G06F 2213/0026 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A processing device, comprising:
a storage unit configured to store multiple input neurons, at least one output neuron, multiple weights, and an instruction of a neural network;
a coarse-grained pruning unit configured to perform coarse-grained pruning on the weights of the neural network to obtain pruned weights and store the pruned weights and position information of one or more target weights into the storage unit, wherein an absolute value of each of the target weights is greater than a positive second given threshold, and the coarse-grained pruning unit is specifically configured to:
select one or more weights from weights of the neural network through a sliding window; and
when the one or more weights meet a preset condition, set the selected weights to 0;
determine position information of the target weights among the pruned weights based on the position of the selected weights that are set to zero;
a coarse-grained selection unit configured to
receive the multiple input neurons, the position information of the target weights, and the target weights and
select one or more input neurons corresponding to the target weights according to the position information of the target weights; and
an operation unit configured to perform training according to the pruned weights, and the selected weights that have been set to 0 in the training process remain 0, perform a neural network operation according to the target weights and the selected input neurons corresponding to the target weights to generate the at least an one output neuron, and transmit the output neuron to the storage unit as an input neuron of a next layer.