DISTRIBUTED DEEP LEARNING SYSTEM

WIPO Patent Application WO/2019/159783

A1

In the present invention, distributed deep learning is performed at a high speed. Learning nodes (2-0)-(2-3) each calculate the gradient of a loss function from an output result obtained by inputting learning data to a learning-target neural network, packetize a calculation result, and transmit the packetized calculation result to a computing interconnect device (1). The computing interconnect device (1) receives the packets transmitted from the learning nodes (2-0)-(2-3), acquires the values of the gradients stored in these packets, calculates the sum of the gradients, packetizes a calculation result, and transmits the packetized calculation result to the learning nodes (2-0)-(2-3). The learning nodes (2-0)-(2-3) receive the packets transmitted from the computing interconnect device (1), and update a configuration parameter of the neural network on the basis of the values stored in these packets.

KATO JUNICHI (JP)
KAWAI KENJI (JP)
NGO HUYCU (JP)
ARIKAWA YUKI (JP)
ITO TSUYOSHI (JP)
SAKAMOTO TAKESHI (JP)

PCT/JP2019/004213

August 22, 2019

February 06, 2019

Click for automatic bibliography generation  

NIPPON TELEGRAPH & TELEPHONE (JP)

G06N3/08

HIDAKA, MASATOSHI ET AL.,: "Development of environment-independent and distributed computable deep convolutional neural network framework by JavaScript", IPSJ SIG TECHNICAL REPORTS: GRAPHICS AND CAD (CG), vol. 2015-CG-161, no. 3, 2015, pages 1 - 9
DEAN, JEFFREY ET AL.: "Large Scale Distributed Deep Networks", PROCEEDINGS OF ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEM 25(NIPS2012), 2012, pages 1 - 9, XP055534679, Retrieved from the Internet [retrieved on 20190423]

YAMAKAWA, Shigeki et al. (JP)

View/Download PDF      PDF Help

Previous Patent: GAS SEPARATION DEVICE, GAS SEPARATION METHOD, AND GAS SEPARATION MEMBRANE

Next Patent: DISTRIBUTED PROCESSING SYSTEM AND DISTRIBUTED PROCESSING METHOD