DISTRIBUTED PROCESSING OF ARTIFICAL NEURAL NETWORK AND DEVICE PERFORMING THE SAME
Disclosed are a distributed processing method of an artificial neural network and a device for performing the same. The distributed processing method according to various embodiments, which is a distributed processing method of a layer calculation job occurring at each layer of the artificial neural...
Gespeichert in:
Hauptverfasser: | , |
---|---|
Format: | Patent |
Sprache: | eng ; kor |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Disclosed are a distributed processing method of an artificial neural network and a device for performing the same. The distributed processing method according to various embodiments, which is a distributed processing method of a layer calculation job occurring at each layer of the artificial neural network, comprises: an operation of generating a layered graph by stacking a network graph corresponding to a physical computing network including a plurality of physical computing nodes for distributed processing of the layer calculation job at a plurality of levels; and an operation of minimizing a job completion time in each network graph based on the job completion time for the layer calculation job and generating a shortest path including a virtual computing node reflecting an input-output relationship of the layer calculation job in the layered graph. According to various embodiments, provided is a technique for allocating computing calculation jobs to the computing nodes according to the shortest path and establishing a path between the computing nodes to perform the computing calculation jobs within a minimum job completion time.
인공 신경망의 분산 처리 방법 및 이를 수행하는 장치가 개시된다. 다양한 실시예에 따른 분산 처리 방법은 인공 신경망의 각 계층에서 발생한 계층 계산 작업을 분산 처리하는 방법에 있어서 상기 계층 계산 작업을 분산 처리할 물리적 컴퓨팅 노드를 복수 개 포함하는 물리적 컴퓨팅 네트워크에 대응하는 네트워크 그래프를 복수 레벨로 적층하여 계층 그래프(layered graph)를 생성하는 동작과 상기 계층 계산 작업에 대한 작업 완료 시간(job completion time)에 기초하여 각 네트워크 그래프에서 상기 작업 완료 시간을 최소화하고 상기 계층 계산 작업의 입출력 관계를 반영하는 가상 컴퓨팅 노드가 포함된 최단 경로를 상기 계층 그래프에서 생성하는 동작을 포함할 수 있다. |
---|