Title:
REGULARIZATION OF NEURAL NETWORK
Document Type and Number:
Japanese Patent JP2018041367
Kind Code:
A
Abstract:
PROBLEM TO BE SOLVED: To solve the problem that application of an existing regularization method is difficult due to the DyBM of a new architecture.SOLUTION: The present invention includes: acquiring an original neural network having a plurality of first-in first-out (FIFO) queues where each FIFO queue is located between a pair of nodes among a plurality of nodes of the original neural network; generating at least one altered neural network, equal to the original neural network, whose length is altered and which has at least one FIFO queue; evaluating the original neural network and each neural network out of at least one altered neural network; and determining, on the basis of the evaluation, which neural network out of the original neural network and at least one altered neural network is most accurate; whereby a neural network can be regularized.SELECTED DRAWING: Figure 1
Inventors:
SAKYASINGHA DASGUPTA
OSOGAMI TAKAYUKI
OSOGAMI TAKAYUKI
Application Number:
JP2016176336A
Publication Date:
March 15, 2018
Filing Date:
September 09, 2016
Export Citation:
Assignee:
INT BUSINESS MASCHINES CORPORATION
International Classes:
G06N3/08; G06N3/04
Domestic Patent References:
JP2016071697A | 2016-05-09 | |||
JP2015011510A | 2015-01-19 | |||
JP2013143031A | 2013-07-22 | |||
JP2011065361A | 2011-03-31 |
Foreign References:
US20150106316A1 | 2015-04-16 | |||
US20150006444A1 | 2015-01-01 |
Attorney, Agent or Firm:
Takeshi Ueno
Tasaichi Tanae
Akashi Hideya
Tasaichi Tanae
Akashi Hideya