Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
CALCULATION CIRCUIT, CALCULATION METHOD, AND PROGRAM
Document Type and Number:
WIPO Patent Application WO/2019/053835
Kind Code:
A1
Abstract:
In an embodiment, provided is a calculation method of a convolution layer in a convolution neural network. The calculation method is provided with: a step for generating a coefficient matrix (A) by converting a kernel used in the convolution layer so as to correspond to an input vector (x) obtained by deploying, in a row, a feature map input to the convolution layer; a step for searching for non-zero elements included in the coefficient matrix; a step for taking, as a processing unit, a multiplication of each non-zero element included in the coefficient matrix and a corresponding element of the input vector, and allocating the multiplication to each of a plurality of calculators (CL) which can perform parallel processing, so as to equalize the number of processing units in each calculator; and a step for sequentially executing, by each calculator, the multiplications allocated thereto and integrating the multiplication results to a corresponding element of an output vector (f).

Inventors:
TANAKA SUSUMU (JP)
MORI MASASHI (JP)
HASHIMOTO KAZUSHIGE (JP)
Application Number:
PCT/JP2017/033225
Publication Date:
March 21, 2019
Filing Date:
September 14, 2017
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
MITSUBISHI ELECTRIC CORP (JP)
International Classes:
G06F17/10; G06N3/063
Foreign References:
US20160358069A12016-12-08
Other References:
OSODA, TSUTOMA ET AL.: "Vectorization and parallelization technique of block ILU preconditioning for unstructural problems", TRANSACTIONS OF THE INFORMATION PROCESSING SOCIETY OF JAPAN, vol. 41, no. SIG8 (HPS2), November 2000 (2000-11-01), pages 92 - 100
Attorney, Agent or Firm:
FUKAMI PATENT OFFICE, P.C. (JP)
Download PDF: