SELECTIVE INFERENCE METHOD IN A ARTIFICIAL NEURAL NETWORK DEVICE

A selective inference method in an artificial neural network is disclosed. The selective inference method includes: a step of allowing an intermediate output layer to receive an operation result wherein the operation result is generated by operation from an arbitrary intermediate layer among a plura...

Full description

Saved in:
Bibliographic Details
Main Authors LEE, MUN SEOB, LEE, HYUN YONG, LEE, SEI HYOUNG, KIM, YOUNG SUN, KIM, NACK WOO, LEE, BYUNG TAK
Format Patent
LanguageEnglish
Korean
Published 08.06.2018
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:A selective inference method in an artificial neural network is disclosed. The selective inference method includes: a step of allowing an intermediate output layer to receive an operation result wherein the operation result is generated by operation from an arbitrary intermediate layer among a plurality of intermediate layers to the arbitrary intermediate layer with respect to data input through an input layer and operating an inference result value using the received operation result; and a step of allowing an inference adjuster to determine whether an inference result value received from the intermediate output layer satisfies a certain criterion, and decide whether to end inference by operation to the arbitrary intermediate layer or continue the inference by performing operation to the output layer from an intermediate layer after the arbitrary intermediate layer according to a determination result. It is possible to reduce time required for the inference of a deep artificial neural network. 인공신경망에서의 선택적 추론 방법이 개시된다. 이 선택적 추론 방법은, 중간 출력층이, 상기 다수의 중간층 중에서 임의의 중간층으로부터, 상기 입력층을 통해 입력된 데이터에 대해 상기 임의의 중간층까지 연산한 연산결과를 전달받고, 전달받은 상기 연산결과를 이용하여 추론 결과값을 연산하는 단계; 및 추론 조절자가, 상기 중간 출력층으로부터 전달받은 상기 추론 결과값이 일정 기준을 만족하는지 판단하고, 판단 결과에 따라, 상기 임의의 중간층까지의 연산으로 추론을 종료할지 아니면 상기 임의의 중간층 이후의 중간층부터 상기 출력층까지의 연산을 수행하여 상기 추론을 계속 진행할지를 결정하는 단계를 포함한다.
Bibliography:Application Number: KR20160161800