Title:
MODEL LEARNING DEVICE, METHOD THEREFOR, AND PROGRAM
Document Type and Number:
WIPO Patent Application WO/2018/051841
Kind Code:
A1
Abstract:
The present invention comprises: an initial value setting unit that, by using a parameter for a first model which has been learned and which includes a neutral network, sets a parameter for a second model which includes a neutral network having a structure similar to that of the first model; a first output probability distribution calculating unit that, by using the first model and a feature amount for learning, calculates a first output probability distribution including distribution of output probabilities of units in an output layer; a second output probability distribution calculating unit that, by using the second model and the feature amount for learning, calculates a second output probability distribution including distribution of the output probabilities of the units in the output layer; and a correction model updating unit that obtains a weighted sum of a second loss function calculated from the second output probability distribution and from information about a correct answer and cross entropy between the first output probability distribution and the second output probability distribution, and that updates the parameter for the second model such that the weighted sum is decreased.
Inventors:
MASATAKI HIROKAZU (JP)
ASAMI TAICHI (JP)
NAKAMURA TAKASHI (JP)
MASUMURA RYO (JP)
ASAMI TAICHI (JP)
NAKAMURA TAKASHI (JP)
MASUMURA RYO (JP)
Application Number:
PCT/JP2017/031909
Publication Date:
March 22, 2018
Filing Date:
September 05, 2017
Export Citation:
Assignee:
NIPPON TELEGRAPH & TELEPHONE (JP)
International Classes:
G10L15/16; G06N3/08; G06N99/00; G10L15/06; G10L15/065
Domestic Patent References:
| WO2016167779A1 | 2016-10-20 |
Foreign References:
| US20150161995A1 | 2015-06-11 | |||
| US20170011738A1 | 2017-01-12 |
Other References:
CUI, YIXIN ET AL.: "Extreme learning machine based on cross entropy", MACHINE LEARNING AND CYBERNETICS, vol. 2, 13 June 2016 (2016-06-13), pages 1066 - 1071, XP033073942, ISBN: 978-1-5090-0390-7
Attorney, Agent or Firm:
NAKAO, Naoki et al. (JP)
Download PDF: