METHOD AND APPARATUS OF ARTIFICIAL NEURAL NETWORK QUANTIZATION

A quantization method of an artificial neural network simultaneously providing high accuracy and low computational complexity according to an exemplary embodiment of the present disclosure can comprise the steps of: dividing input distribution of an artificial neural network into a plurality of segm...

Full description

Saved in:
Bibliographic Details
Main Authors YIM HAN YOUNG, LIM JONG HAN, KIM BYEOUNG SU, SUNG NAK WOO, KANG IN YUP, HA SANG HYUCK, KIM DO YUN
Format Patent
LanguageEnglish
Korean
Published 03.04.2019
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:A quantization method of an artificial neural network simultaneously providing high accuracy and low computational complexity according to an exemplary embodiment of the present disclosure can comprise the steps of: dividing input distribution of an artificial neural network into a plurality of segments; generating an approximate density function by approximating each of the segments; calculating at least one quantization error corresponding to at least one step size based on the approximate density function; and determining a final step size based on at least one quantization error. 본 개시의 예시적 실시예에 따른 인공 신경망의 양자화 방법은, 인공 신경망의 입력 분포를 복수의 세그먼트들로 분할하는 단계, 복수의 세그먼트들 각각을 근사화함으로써 근사 밀도 함수를 생성하는 단계, 근사 밀도 함수에 기초하여, 적어도 하나의 스텝 사이즈에 대응하는 적어도 하나의 양자화 오차를 계산하는 단계, 및 적어도 하나의 양자화 오차에 기초하여, 최종 스텝 사이즈를 결정하는 단계를 포함할 수 있다.
Bibliography:Application Number: KR20170123658