US 11,811,427 B2
Information processing apparatus, method of processing information, and non-transitory computer-readable storage medium for storing information processing program
Yasufumi Sakai, Fuchu (JP); and Sosaku Moriki, Fukuoka (JP)
Assigned to FUJITSU LIMITED, Kawasaki (JP)
Filed by FUJITSU LIMITED, Kawasaki (JP)
Filed on Jul. 16, 2020, as Appl. No. 16/930,356.
Claims priority of application No. 2019-167608 (JP), filed on Sep. 13, 2019.
Prior Publication US 2021/0081783 A1, Mar. 18, 2021
Int. Cl. H03M 7/30 (2006.01); G06N 3/04 (2023.01); G06N 3/08 (2023.01); G06F 18/21 (2023.01)
CPC H03M 7/3059 (2013.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01); G06F 18/21 (2023.01)] 6 Claims
OG exemplary drawing
 
1. An information processing apparatus of performing quantization processing on quantization target data using a fixed-point type variable, the information processing apparatus comprising:
an accelerator circuit configured to execute calculation of a neural network; and
a quantization processing circuit coupled to the accelerator circuit, the quantization processing circuit including:
a statistical processing circuit configured to obtain a distribution of appearance frequencies of a plurality of variable elements included in the quantization target data, each of the plurality of variable elements being a floating-point type variable; and
a quantization position setting circuit configured to:
align a most significant bit (MSB) position of a quantization position to be used for conversion from the floating-point type variable to a fixed-point type variable, to a variable element smaller than a variable element of a maximum value among the plurality of variable elements based on the distribution of the appearance frequencies of the plurality of variable elements, the aligning of the quantization position including: adjusting at least the maximum value among the plurality of variable elements to fall outside the quantization position, and adjusting the minimum value among the plurality of variable elements to fall inside the quantization position; and
convert, for each variable element of the plurality of variable elements, the variable element into a respective fixed-point type variable by using the aligned quantization position, to perform machine-learning by causing the accelerator circuit to execute the calculation of the neural network using the respective fixed-point type variable converted from the variable element.