US 12,282,842 B2
Neural network processing apparatus, neural network processing method, and neural network processing program
Hiroyuki Tokunaga, Tokyo (JP)
Assigned to MAXELL, LTD., Kyoto (JP)
Appl. No. 17/292,661
Filed by MAXELL, LTD., Kyoto (JP)
PCT Filed Sep. 10, 2019, PCT No. PCT/JP2019/035493
§ 371(c)(1), (2) Date May 10, 2021,
PCT Pub. No. WO2020/100393, PCT Pub. Date May 22, 2020.
Claims priority of application No. 2018-214570 (JP), filed on Nov. 15, 2018.
Prior Publication US 2022/0004857 A1, Jan. 6, 2022
Int. Cl. G06N 3/06 (2006.01); G06F 9/50 (2006.01); G06N 3/063 (2023.01)
CPC G06N 3/063 (2013.01) [G06F 9/5016 (2013.01)] 2 Claims
OG exemplary drawing
 
1. A neural network processing apparatus comprising:
a first memory configured to store a value given to a neural network;
a second memory configured to store a table in which the value given to the neural network and a binary signal corresponding to the value are associated with each other;
a binary signal acquisition circuit configured to acquire the binary signal corresponding to the value stored in the first memory by referring to the table; and
a neural network operation circuit configured to perform an operation of the neural network based on the binary signal acquired by the binary signal acquisition circuit,
wherein
the binary signal has an accuracy complying with a predetermined operation accuracy of the neural network, and the second memory stores a plurality of tables,
the binary signal acquisition circuit acquires the binary signal corresponding to the value stored in the first memory by referring to one table selected from the plurality of tables;
the neural network processing apparatus further comprising:
a table selector configured to select the one table from the plurality of tables based on a reference concerning the predetermined accuracy in accordance with the value stored in the first memory, wherein the binary signals stored in the plurality of tables have accuracies different from each other,
a determination circuit configured to determine whether the value stored in the first memory is a value representing specific information; and
a converter configured to convert, based on a determination result by the determination circuit, a data type and a size of the value stored in the first memory,
wherein the plurality of tables are provided in correspondence with at least the data types and the sizes of the value given to the neural network,
the table selector selects, from the plurality of tables, a table that matches the data type and the size of the value converted by the converter, and
the binary signal acquisition circuit acquires the binary signal corresponding to the value converted by the converter by referring to the table selected by the table selector.