Title:
LEARNING DEVICE AND LEARNING METHOD
Document Type and Number:
Japanese Patent JP2019212171
Kind Code:
A
Abstract:
To provide a learning device and a learning method with which it is possible to create a histogram with respect to the feature quantity and gradient information that are inputted in parallel and cut down the circuit scale of a memory that holds information pertaining to the histogram.SOLUTION: Provided is a learning device that performs learning by gradient boosting, comprising: a plurality of data storage units for storing learning data including a feature quantity and gradient information; a gradient output unit provided in plurality in correspondence to each data storage unit, for inputting a feature quantity and gradient information from the data storage unit, and outputting gradient information that corresponds to the feature quantity from an output port that corresponds to each value of the inputted feature quantity; an addition unit for adding gradient information that corresponds to the same value of feature quantity outputted from each gradient output unit and outputting the added value of gradient information that corresponds to each value; and a histogram storage unit for storing a histogram in which the added value of gradient information that corresponds to each value of feature quantity outputted from the added value is integrated for each bin, the bin representing each value.SELECTED DRAWING: Figure 22
Inventors:
TANAKA TAKUYA
KASAHARA RYOSUKE
KASAHARA RYOSUKE
Application Number:
JP2018109735A
Publication Date:
December 12, 2019
Filing Date:
June 07, 2018
Export Citation:
Assignee:
RICOH CO LTD
International Classes:
G06N20/00
Domestic Patent References:
| JP2013008270A | 2013-01-10 |
Foreign References:
| WO2016122636A1 | 2016-08-04 |
Other References:
GUOLIN KE, ET AL.: "LightGBM: A Highly Efficient Gradient Boosting Decision Tree", NIPS'17: PROCEEDINGS OF THE 31ST INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS, JPN6022016978, December 2017 (2017-12-01), pages 3149 - 3157, ISSN: 0004765843
Attorney, Agent or Firm:
Hiroaki Sakai