Feature Extraction Method and System through Attentive Convolution Operation

Provided are a method and system for extracting a feature through an attentive convolution operation. In the convolution operation that extracts a feature of an input tensor, a query tensor and key tensor are created using a query matrix and key matrix which are capable of learning before the convol...

Full description

Saved in:
Bibliographic Details
Main Authors KIM SAN, SHIN SA IM, JUNG HYE DONG
Format Patent
LanguageEnglish
Korean
Published 10.05.2022
Subjects
CALCULATING
COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
COMPUTING
COUNTING
ELECTRIC DIGITAL DATA PROCESSING
PHYSICS
Online AccessGet full text

Cover

More Information
Summary:Provided are a method and system for extracting a feature through an attentive convolution operation. In the convolution operation that extracts a feature of an input tensor, a query tensor and key tensor are created using a query matrix and key matrix which are capable of learning before the convolution operation, and attention is created using the tensors. Using the attention, a filter value is adjusted according to a value of the input tensor for each unit operation, and the output is created with the adjusted filter value. 어텐티브 합성곱 연산을 통한 특징 추출 방법 및 시스템이 제공된다. 입력 텐서의 특징을 추출하는 합성곱 연산에서 합성곱 연산 전 학습이 가능한 쿼리 행렬과 키 행렬을 이용하여 쿼리 텐서와 키 텐서를 생성하여 이러한 텐서들을 이용하여 어텐션을 만들고, 어텐션을 이용하여 각 단위 연산마다 입력 텐서의 값에 따라 필터 값을 조정하고 조정된 필터값으로 출력을 생성하도록 한다.
Bibliography:Application Number: KR20200144053