DEEP NEURAL NETWORK ACCELERATING DEVICE AND OPERATION METHOD THEREOF
A deep neural network learning acceleration device according to an embodiment of the present application comprises: a calculation part that sequentially performs first and second calculations for a plurality of input data of a subset, according to a mini-batch gradient descent method; a determinatio...
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: | A deep neural network learning acceleration device according to an embodiment of the present application comprises: a calculation part that sequentially performs first and second calculations for a plurality of input data of a subset, according to a mini-batch gradient descent method; a determination part that determines each of the plurality of input data as one among the skip data or the learning data, based on a reliability matrix obtained according to the first calculation; and a control part that controls the calculation part to skip the second calculation for the skip data. Therefore, the present invention is capable of reducing energy consumption.
본 출원의 실시예에 따른 심층 신경망 학습 가속 장치는 미니-배치 경사 하강법에 따라, 서브 세트의 복수의 입력 데이터들에 대한 제1 및 제2 연산을 순차적으로 수행하는 연산부, 상기 제1 연산에 따라 획득되는 신뢰도 행렬에 기초하여, 상기 복수의 입력 데이터들 각각을 스킵 데이터와 학습 데이터 중 어느 하나로 판단하는 판단부 및 상기 스킵 데이터에 대한 상기 제2 연산을 스킵시키도록 상기 연산부를 제어하는 제어부를 포함한다. |
---|