US 11,811,421 B2
Weights safety mechanism in an artificial neural network processor
Guy Kaminitz, Kfar Saba (IL); Roi Seznayov, Raanana (IL); Daniel Chibotero, Kirat Ono (IL); Ori Katz, Tel-Aviv (IL); Nir Engelberg, Tel-Aviv (IL); Yuval Adelstein, Tel-Aviv (IL); Or Danon, Kirat Ono (IL); and Avi Baum, Givat Shmuel (IL)
Assigned to Hailo Technologies Ltd.
Filed by Hailo Technologies Ltd., Tel Aviv (IL)
Filed on Sep. 29, 2020, as Appl. No. 17/035,892.
Prior Publication US 2022/0103186 A1, Mar. 31, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. H03M 13/09 (2006.01); G06F 11/14 (2006.01); H03M 13/15 (2006.01); G06N 3/08 (2023.01); G06N 3/063 (2023.01); G06F 18/22 (2023.01); G06F 18/231 (2023.01)
CPC H03M 13/096 (2013.01) [G06F 11/1476 (2013.01); G06F 18/22 (2023.01); G06F 18/231 (2023.01); G06N 3/063 (2013.01); G06N 3/08 (2013.01); H03M 13/15 (2013.01)] 18 Claims
OG exemplary drawing
 
15. A method of detecting errors in weights used in a neural network processor, the method comprising:
parsing a model of a neural network and determining a CRC block size n for each layer of the neural network;
calculating a CRC checksum value for each block of n weights, said CRC checksum values operative to protect the weights in each block;
populating a memory with blocks of weights including associated precalculated CRC checksums;
configuring a layer control unit (LCU) circuit for each layer of the neural network to skip over precalculated CRC checksums in each block during neural network calculations;
wherein said weights are subsequently read from the memory and used in neural network calculations performed on said processor; and
wherein said CRC checksum values are subsequently read from the memory and compared to CRC checksums calculated over weights read from the memory for each block and used in the neural network calculations.