Title:
CONVOLUTION NEURAL NETWORK
Document Type and Number:
Japanese Patent JP2018200627
Kind Code:
A
Abstract:
To provide a convolution neural network capable of applying, even when a large number of input bars are required for a convolution operation in a convolution layer, a crossbar circuit to the convolution operation.SOLUTION: A crossbar circuit is divided into a first crossbar circuit 61 and a second crossbar circuit 62. The first crossbar circuit 61 performs a convolution operation with a plurality of filters to a part of feature maps among feature maps included in input data. The second crossbar circuit 62 performs a convolution operation with a plurality of filters to another part different from the part of feature maps to be subjected to the convolution operation in the first crossbar circuit 61. The feature map as the convolution operation result of the filters performed by the first crossbar circuit 61 and the feature map as the result of the convolution operation of the filters performed by the second crossbar circuit 62 are merged in a merge layer 63.SELECTED DRAWING: Figure 8
Inventors:
IRINA KATAEVA
Application Number:
JP2017105742A
Publication Date:
December 20, 2018
Filing Date:
May 29, 2017
Export Citation:
Assignee:
DENSO CORP
International Classes:
G06G7/60; G06N3/063
Attorney, Agent or Firm:
Kazuyuki Yahagi
Taihei Nonobe
Takanori Kubo
Taihei Nonobe
Takanori Kubo