Title:
DECONVOLUTIONAL NEURAL NETWORK HARDWARE ACCELERATION METHOD, APPARATUS, AND ELECTRONIC DEVICE
Document Type and Number:
WIPO Patent Application WO/2020/118608
Kind Code:
A1
Abstract:
A deconvolutional neural network hardware acceleration method, an apparatus, and an electronic device, the method comprising: input data for a current network layer is acquired (S1), the input data being a calculation result from a previous network layer stored in a first memory; deconvolutional calculation is performed on the input data on the current network layer to obtain a current calculation result, and the current calculation result is input into a second memory (S2), the current network layer comprising a weight of said layer; on the basis of the current calculation result in the second memory, input data for a next network layer is obtained (S3); deconvolutional calculation is performed on the input data for the next network layer on the next network layer to obtain a current calculation result, and the current calculation result is input into the first memory (S4), the next network layer comprising a weight of said layer; and the above steps are repeated until the last layer of the deconvolutional neural network, and a result is output (S5). By means of the described method, the efficiency of data transmission and usage is improved.
Inventors:
LIU SHUANGLONG (CN)
Application Number:
PCT/CN2018/120861
Publication Date:
June 18, 2020
Filing Date:
December 13, 2018
Export Citation:
Assignee:
SHENZHEN CORERAIN TECH CO LTD (CN)
International Classes:
G06N3/04
Foreign References:
| CN108875915A | 2018-11-23 | |||
| CN108876833A | 2018-11-23 | |||
| CN108765282A | 2018-11-06 | |||
| CN108133265A | 2018-06-08 | |||
| US20180060724A1 | 2018-03-01 |
Attorney, Agent or Firm:
SHENZHEN HYVISION INTELLECTUAL PROPERTY ATTORNEY (CN)
Download PDF: