Title:
NEURAL NETWORK TRAINING METHOD AND DEVICE, COMPUTER SYSTEM AND MOBILE DEVICE
Document Type and Number:
WIPO Patent Application WO/2018/218481
Kind Code:
A1
Abstract:
A neural network training method and device, a computer system and a mobile device. The method comprises: acquiring a first weight set of processing units of a neural network (410); performing three-value processing on each weight in a first weight set to obtain a second weight set (420); generating an output value of the processing unit according to the second weight set and an input value set of the processing unit, wherein the input value set of the processing unit of the next layer of the neural network is taken from the output value of the processing unit of the previous layer (430); and training the weights in the first weight set of each processing unit of the neural network according to an error cost function, wherein the error cost function comprises an error term and a structural sparse term, the structural sparse term causing all weights in the first weight set of some of the processing units of the neural network to be zero (440). The described method may reduce the resource requirements of the neural network.
Inventors:
PAN WEI (CN)
CUI JIAN (CN)
LIN XIAOFAN (CN)
ZHAO CONG (CN)
CUI JIAN (CN)
LIN XIAOFAN (CN)
ZHAO CONG (CN)
Application Number:
PCT/CN2017/086547
Publication Date:
December 06, 2018
Filing Date:
May 31, 2017
Export Citation:
Assignee:
SZ DJI TECHNOLOGY CO LTD (CN)
International Classes:
G06N3/08
Foreign References:
| CN106548127A | 2017-03-29 | |||
| CN104091081A | 2014-10-08 | |||
| CN105701480A | 2016-06-22 | |||
| US9552549B1 | 2017-01-24 |
Attorney, Agent or Firm:
LONGSUN LEAD IP LTD. (CN)
Download PDF: