Title:
NEURAL NETWORK PROCESSOR, NEURAL NETWORK PROCESSING METHOD, AND PROGRAM
Document Type and Number:
WIPO Patent Application WO/2020/026475
Kind Code:
A1
Abstract:
The present invention realizes a neural network processor which has a high-performance compact model mountable onto a low-specification device such as a built-in device or a mobile device without needing to learn. A neural network processor (100) can carry out a convolution layer process and a fully-connected layer process by commonizing portions of similar processes in the convolution layer process and the fully-connected layer process, and executing a process obtained by combining processes of a norm mode and an inner product calculation mode. Thus, the neural network processor (100) can carry out a multivalued neural network process at a high speed while curbing increases in hardware size.
Inventors:
MATSUMOTO MAHITO (JP)
ISHIO KOJI (JP)
ISHIO KOJI (JP)
Application Number:
PCT/JP2019/005465
Publication Date:
February 06, 2020
Filing Date:
February 15, 2019
Export Citation:
Assignee:
MEGACHIPS CORP (JP)
A SUM TECH LLC (JP)
A SUM TECH LLC (JP)
International Classes:
G06N3/063
Other References:
KAMIYA, RYUJI ET AL.: "Speed-up and Model Compression of Identification Calculation by Binarized-CNN", IEICE TECHNICAL REPORT, vol. 116, no. 366, 8 December 2016 (2016-12-08), pages 47 - 52
Attorney, Agent or Firm:
NAKANISHI Ken et al. (JP)
Download PDF: