Title:
DATA AMOUNT COMPRESSION METHOD, DEVICE, PROGRAM AND IC CHIP
Document Type and Number:
WIPO Patent Application WO/2019/026702
Kind Code:
A1
Abstract:
[Problem] To compress a data amount of a learning-finished model while minimizing the degradation in prediction accuracy. [Solution] Provided is a data amount compression method which compresses a data amount that corresponds to a learning-finished model obtained by allowing a prescribed data group to be learnt by a learning model which has a tree structure configured by hierarchically disposing a plurality of nodes respectively associated with hierarchically divided state spaces. An error amount, which is generated in the learning process and corresponds to prediction accuracy, is associated with each node of the learning-finished model. The data amount compression method is provided with: a reading step for reading the error amount associated with each of the nodes; and a node removing step for removing a portion of nodes of the learning-finished model on the basis of the error amount read in the reading step, and thereby compressing the data amount corresponding to the learning-finished model.
Inventors:
KIM CHYON HAE (JP)
IDESAWA JUNICHI (JP)
SUGAWARA SHIMON (JP)
IDESAWA JUNICHI (JP)
SUGAWARA SHIMON (JP)
Application Number:
PCT/JP2018/027773
Publication Date:
February 07, 2019
Filing Date:
July 24, 2018
Export Citation:
Assignee:
AISING LTD (JP)
International Classes:
G06N99/00
Foreign References:
JP2016173686A | 2016-09-29 | |||
JP2017126890A | 2017-07-20 | |||
JP2016173686A | 2016-09-29 |
Other References:
KIM, CHYON-HAE ET AL.: "Error based prediction algorithm in dynamics learning tree", IPSJ SIG TECHNICAL REPORTS: MATHEMATICAL MODELING AND PROBLEM SOLVING (MPS, vol. 2017-MPS-12, no. 25, 20 February 2017 (2017-02-20), pages 1 - 5, XP009515865, ISSN: 2188-8833
IDESAWA, JUNICHI ET AL.: "Japan 's first'' ''World's first'' machine-control AI DeepBinaryTree: DBT ''machine learning algorithm that can be embedded to lightweight edge device", LECTURE MATERIALS, 27 April 2018 (2018-04-27), pages 1 - 28, XP009515772, Retrieved from the Internet [retrieved on 20180906]
See also references of EP 3663991A4
IDESAWA, JUNICHI ET AL.: "Japan 's first'' ''World's first'' machine-control AI DeepBinaryTree: DBT ''machine learning algorithm that can be embedded to lightweight edge device", LECTURE MATERIALS, 27 April 2018 (2018-04-27), pages 1 - 28, XP009515772, Retrieved from the Internet
See also references of EP 3663991A4
Attorney, Agent or Firm:
IIZUKA Shin-ichi et al. (JP)
Download PDF: