Title:
CONVOLUTIONAL NEURAL NETWORK
Document Type and Number:
WIPO Patent Application WO/2018/220957
Kind Code:
A1
Abstract:
A neural network (20) is provided with a plurality of convolution layers (21-27) and a merging layer (63). One convolution layer has a crossbar circuit (44) that has a plurality of input bars (50), a plurality of output bars (51, 52), and a plurality of weight assignment elements (53) which assign weights. The crossbar circuit (44) carries out convolution calculation in an analog region by assigning weights to input signals and adding the input signals together on each of the output bars. Input data includes a plurality of feature maps. The crossbar circuit (44) has a first crossbar circuit (61) that carries out convolution calculation for a portion of the feature maps, and a second crossbar circuit (62) that carries out convolution calculation for another portion of the feature maps. The merging layer (63) merges the calculation results of the first and second crossbar circuits.
Inventors:
KATAEVA IRINA (JP)
Application Number:
PCT/JP2018/011272
Publication Date:
December 06, 2018
Filing Date:
March 22, 2018
Export Citation:
Assignee:
DENSO CORP (JP)
International Classes:
G06N3/063; G06G7/60
Foreign References:
| US20180018559A1 | 2018-01-18 | |||
| US9646243B1 | 2017-05-09 |
Other References:
YAKOPCIC, CHRIS ET AL.: "Memristor crossbar deep network implementation based on a convolutional neural network", IEEE 2016 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN), 24 July 2016 (2016-07-24) - 29 July 2016 (2016-07-29), XP032992268, ISSN: 2161-4407, DOI: 10.1109/IJCNN.2016.7727302
Attorney, Agent or Firm:
JIN Shunji (JP)
Download PDF:
Previous Patent: CONTROLLER FOR VEHICLES AND METHOD OF REWRITING PROGRAM
Next Patent: LIGHT-RECEIVING MODULE FOR ENCODER, AND ENCODER
Next Patent: LIGHT-RECEIVING MODULE FOR ENCODER, AND ENCODER