METHOD FOR NEURAL NETWORK QUANTIZATION USING RANDOMIZED SCALE

Provided is a neural network quantization technique through arbitrary scale transformation, which includes the steps of: quantizing input data and parameters through a quantization module; converting quantized input data and quantized parameters into quantized output data through a quantized operati...

Full description

Saved in:
Bibliographic Details
Main Authors DAO THANH TUAN, KIM HEEHOON, JUNG WOOKEUN, LEE JAEJIN
Format Patent
LanguageEnglish
Korean
Published 04.08.2022
Subjects
CALCULATING
COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
COMPUTING
COUNTING
PHYSICS
Online AccessGet full text

Cover

More Information
Summary:Provided is a neural network quantization technique through arbitrary scale transformation, which includes the steps of: quantizing input data and parameters through a quantization module; converting quantized input data and quantized parameters into quantized output data through a quantized operation module; and re-converting the transformed quantized output data to before quantization through an inverse quantization module. According to the present invention, since there is an effect of filtering out exceptional input values, learning accuracy increases. 임의 스케일 변환을 통한 뉴럴 네트워크 양자화 기법에 있어서, 입력 데이터 및 파라미터를 양자화 모듈을 통해 양자화하는 단계, 양자화된 입력 데이터 및 양자화된 파라미터를 양자화된 연산 모듈을 통해 양자화된 출력 데이터로 변환하는 단계 및 변환된 상기 양자화된 출력 데이터를 역 양자화 모듈을 통해 양자화되기 전으로 재 변환하는 단계를 포함한다.
Bibliography:Application Number: KR20210012632