US 12,333,437 B2
Arithmetic apparatus and arithmetic method
Daisuke Miyashita, Kawasaki (JP); and Asuka Maki, Kawasaki (JP)
Assigned to Kioxia Corporation, Tokyo (JP)
Filed by Kioxia Corporation, Tokyo (JP)
Filed on Mar. 9, 2021, as Appl. No. 17/195,865.
Claims priority of application No. 2020-151296 (JP), filed on Sep. 9, 2020.
Prior Publication US 2022/0076122 A1, Mar. 10, 2022
Int. Cl. G06N 3/08 (2023.01); G06F 3/06 (2006.01); G06N 3/04 (2023.01); G06N 3/082 (2023.01)
CPC G06N 3/082 (2013.01) [G06F 3/0604 (2013.01); G06F 3/0658 (2013.01); G06F 3/0685 (2013.01); G06N 3/04 (2013.01)] 14 Claims
OG exemplary drawing
 
1. An arithmetic apparatus comprising:
a non-volatile first memory configured to store a model to be trained;
a volatile second memory having a smaller storage capacity than the first memory; and
a processor configured to:
execute learning processing that updates a first parameter of the model based on a loss value obtained by inputting training data into the model stored in the first memory;
perform processing by using the first and second memories while suppressing a number of times of rewriting of the first memory, for each of all layers of neural network constituting the model, the processing including:
reading the first parameter of the model stored in the first memory;
calculating a second parameter in which cumulative update information stored in the second memory is reflected in the first parameter; and
storing a difference between a third parameter obtained by updating the second parameter and the first parameter, in the second memory as the cumulative update information;
determine whether the processing is performed on all the layers of the neural network constituting the model; and
reflect, in response to the processing being performed on all the layers, the cumulative update information stored in the second memory in the model stored in the first memory to update the first parameter stored in the first memory.