Title:
LEARNING DEVICE, LEARNING METHOD, AND LEARNING PROGRAM
Document Type and Number:
WIPO Patent Application WO/2017/183587
Kind Code:
A1
Abstract:
A learning device (10) for practicing learning using a stochastic gradient descent method in machine learning, wherein the device has: a gradient calculation unit (11) for calculating a primary gradient in the stochastic gradient descent method; a statistical quantity calculation unit (12) for calculating the statistical quantity of the primary gradient; an initialization bias elimination unit (13) for eliminating, from the statistical quantity of the primary gradient calculated by the statistical quantity calculation unit (12), an initialization bias that is used as a weight when the statistical quantity calculation unit (12) calculates the statistical quantity of the primary gradient; a learning rate adjustment unit (14) for adjusting a learning rate by dividing the learning rate by a standard deviation of the primary gradient on the basis of the statistical quantity of the primary gradient; and a parameter update unit (15) for updating the parameter of a learning model using the learning rate adjusted by the learning rate adjustment unit (14).
Inventors:
IDA YASUTOSHI (JP)
FUJIWARA YASUHIRO (JP)
FUJIWARA YASUHIRO (JP)
Application Number:
PCT/JP2017/015337
Publication Date:
October 26, 2017
Filing Date:
April 14, 2017
Export Citation:
Assignee:
NIPPON TELEGRAPH & TELEPHONE (JP)
International Classes:
G06N99/00
Other References:
KINGMA, P. ET AL., A METHOD FOR STOCHASTIC OPTIMIZATION, 22 December 2014 (2014-12-22), XP055425253, Retrieved from the Internet [retrieved on 20170630]
Attorney, Agent or Firm:
SAKAI INTERNATIONAL PATENT OFFICE (JP)
Download PDF: