Processing

Please wait...

Settings

Settings

Goto Application

1. WO2020116745 - CONVOLUTION METHOD FOR HIGH-SPEED DEEP LEARNING

Publication Number WO/2020/116745
Publication Date 11.06.2020
International Application No. PCT/KR2019/010624
International Filing Date 21.08.2019
IPC
G06N 3/08 2006.01
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
08Learning methods
G06F 17/15 2006.01
GPHYSICS
06COMPUTING; CALCULATING OR COUNTING
FELECTRIC DIGITAL DATA PROCESSING
17Digital computing or data processing equipment or methods, specially adapted for specific functions
10Complex mathematical operations
15Correlation function computation
CPC
G06F 17/153
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
FELECTRIC DIGITAL DATA PROCESSING
17Digital computing or data processing equipment or methods, specially adapted for specific functions
10Complex mathematical operations
15Correlation function computation ; including computation of convolution operations
153Multidimensional correlation or convolution
G06N 3/08
GPHYSICS
06COMPUTING; CALCULATING; COUNTING
NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
3Computer systems based on biological models
02using neural network models
08Learning methods
Applicants
  • 아주대학교 산학협력단 AJOU UNIVERSITY INDUSTRY-ACADEMIC COOPERATION FOUNDATION [KR]/[KR]
Inventors
  • 선우명훈 SUNWOO, Myung Hoon
  • 김태선 KIM, Tae Sun
Agents
  • 심경식 SHIM, Kyoung-Shik
  • 홍성욱 HONG, Sung-Wook
Priority Data
10-2018-015677707.12.2018KR
Publication Language Korean (KO)
Filing Language Korean (KO)
Designated States
Title
(EN) CONVOLUTION METHOD FOR HIGH-SPEED DEEP LEARNING
(FR) PROCÉDÉ DE CONVOLUTION DESTINÉ À UN APPRENTISSAGE PROFOND À GRANDE VITESSE
(KO) 고속 딥러닝을 위한 컨볼루션 방법
Abstract
(EN)
A convolution method for high-speed deep learning according to the present invention comprises: (a) a step in which a feature map reception unit of a convolution system receives a feature map comprising N channels; (b) a step in which a main control unit of the convolution system selects a “0th” channel from the feature map comprising the N channels; (c) a step in which the main control unit confirms “0,0” xy coordinates from the feature map of the “0th” channel; (d) a coarse step in which a convolution calculation unit of the convolution system performs convolution calculation and rectified linear unit (ReLU) calculation while proceeding by twos in the horizontal direction and vertical direction in the feature map; (e) a step in which a channel switching unit of the convolution system switches to a subsequent channel when the coarse step for the feature map of the “0th” channel is complete; (f) a step in which the main control unit determines whether the switched channel is greater or less than N; and (g) a step in which, if the switched channel in step (f) is greater than N, the main control unit determines the convolution calculation for all channels to have been complete, and outputs the feature map by means of a feature map output unit. Therefore, the convolution calculation which occupies most of a convolution neural network is shortened, and thus the inference speed in deep learning can be increased.
(FR)
Selon la présente invention, un procédé de convolution destiné à un apprentissage profond à grande vitesse comprend : (a) une étape dans laquelle une unité de réception de carte de caractéristiques d'un système de convolution reçoit une carte de caractéristiques comprenant N canaux ; (b) une étape dans laquelle une unité de commande principale du système de convolution sélectionne un « 0 ème » canal dans la carte de caractéristiques comprenant les N canaux ; (c) une étape dans laquelle l'unité de commande principale confirme des coordonnées xy « 0,0 » à partir de la carte de caractéristiques du « 0 ème » canal ; (d) une étape approximative dans laquelle une unité de calcul de convolution du système de convolution effectue un calcul de convolution et un calcul d'unité linéaire rectifiée (ReLU) tout en procédant par deux dans la direction horizontale et dans la direction verticale dans la carte de caractéristiques ; (e) une étape dans laquelle une unité de commutation de canal du système de convolution commute vers un canal suivant lorsque l'étape approximative destinée à la carte de caractéristiques du « 0 ème » canal est terminée ; (f) une étape dans laquelle l'unité de commande principale détermine si le canal commuté est supérieur ou inférieur à N ; et (g) une étape dans laquelle, si le canal commuté à l'étape (f) est supérieur à N, l'unité de commande principale détermine le calcul de convolution de tous les canaux qui ont été terminés, et produit la carte de caractéristiques au moyen d'une unité de sortie de carte de caractéristiques. Par conséquent, le calcul de convolution qui occupe l'essentiel d'un réseau neuronal à convolution est raccourci, d'où la vitesse d'inférence en apprentissage profond peut être augmentée.
(KO)
본 발명에 따른 고속 딥러닝을 위한 컨볼루션 방법은 (a) 상기 컨볼루션 시스템의 피쳐 맵 수신부가 N개의 채널로 구성된 피쳐 맵을 수신하는 단계; (b) 상기 컨볼루션 시스템의 주제어부가 N개의 채널로 구성된 피쳐 맵에서 `0`번째 채널을 선택하는 단계; (c) 상기 주제어부가 `0`번째 채널의 피쳐 맵에서 x, y 좌표가 `0`인 좌표를 확인하는 단계; (d) 상기 컨볼루션 시스템의 컨볼루션 계산부가 상기 피쳐 맵에 대해 가로방향 및 세로방향으로 2만큼 이동해가며 컨볼루션 연산과 ReLU(Rectified Liner Unit)연산을 수행하는 코스(coarse)단계; (e) 상기 컨볼루션 시스템의 채널 변경부가 `0`번째 채널의 피쳐 맵에 대해 코스(coarse)단계가 완료되면 다음 채널로 채널을 변경하는 단계; (f) 상기 주제어부가 변경된 채널이 N보다 큰지 작은지 판단하는 단계; 및 (g) 상기 주제어부가 상기 (f)단계에서 변경된 채널이 N보다 큰 경우 모든 채널에 대한 컨볼루션 연산이 완료된 것으로 판단하고 피쳐 맵 출력부를 통해 피쳐맵을 출력하는 단계;를 포함하여 컨볼루션 신경망에서 대부분을 차지하는 컨볼루션 연산을 줄여 딥러닝에서의 추론 속도를 높일 수 있는 효과가 있다.
Also published as
Latest bibliographic data on file with the International Bureau