Title:
INFORMATION PROCESSING DEVICE, INFORMATION PROCESSING METHOD, AND COMPUTER-READABLE STORAGE MEDIUM
Document Type and Number:
WIPO Patent Application WO/2018/180263
Kind Code:
A1
Abstract:
Provided is an information processing device, comprising: a means which sets, as a plurality of weights of a second NN, a plurality of weights of a selected layer which has been selected from among a plurality of layers of a first NN; a means which classifies each of the plurality of weights of the selected layer into a first group or a second group; a means which assesses a first gradient for each of the weights of the plurality of layers of the first NN on the basis of first teaching data; a second assessment means which, on the basis of second teaching data, assesses a second gradient for the weights, from among the plurality of weights of the second NN, which belong to the first group; and an update means which, for the weights which belong to the first group among the plurality of weights of the selected layer, performs update on the basis of the first gradient and the second gradient, and for the weights which belong to the second group among the plurality of weights of the selected layer, and for the weights of the layers other than the selected layer among the plurality of layers of the first NN, performs update on the basis of the first gradient.
Inventors:
NAGAI YUKI (JP)
KOBAYASHI TATSUYA (JP)
KOBAYASHI TATSUYA (JP)
Application Number:
PCT/JP2018/008445
Publication Date:
October 04, 2018
Filing Date:
March 06, 2018
Export Citation:
Assignee:
KDDI CORP (JP)
International Classes:
G06N3/02; G06F21/16; G06N99/00
Domestic Patent References:
| WO2016043734A1 | 2016-03-24 |
Foreign References:
| JP2017071874A | 2017-04-13 |
Other References:
YANGQING JIAEVAN SHELHAMERJEFF DONAHUESERGEY KARAYEVJONATHAN LONGROSS GIRSHICKSERGIO GUADARRAMATREVOR DARRELL: "Proceedings of the 22nd ACM international conference on Multimedia", ACM, article "Caffe: Convolutional architecture for fast feature embedding", pages: 675 - 678
Attorney, Agent or Firm:
OHTSUKA, Yasunori et al. (JP)
Download PDF: