CROSSBAR ARCHITECTURE-BASED PRUNING METHOD AND SYSTEM

Title:

CROSSBAR ARCHITECTURE-BASED PRUNING METHOD AND SYSTEM

Document Type and Number:

WIPO Patent Application WO/2021/213010

Kind Code:

A1

Abstract:

The present invention relates to a Crossbar architecture-based pruning method and system. The method comprises: carrying out structured pruning without considering architecture factors on a weight matrix to be processed of a current layer of a neural network, to obtain a first weight matrix, the weight matrix being a weight matrix of the neural network; determining a pruning solution according to the size of the first weight matrix and the size of the crossbar, the pruning solution comprising block-based structured pruning, crossbar array row pruning and crossbar array column pruning; and pruning the first weight matrix according to the pruning solution to obtain a second weight matrix, wherein the second weight matrix is used for being mapped to a crossbar array in an accelerator, and the neural network is accelerated by means of the accelerator. The present invention can reduce the resource occupation of a Crossbar and reduce the waste of hardware resources.

More Like This:

JP2022515722	Image segmentation methods and devices, electronic devices and storage media
WO/2022/051855	METHOD AND SYSTEM FOR TRAINING A NEURAL NETWORK MODEL USING GRADUAL KNOWLEDGE DISTILLATION
WO/2021/040699	FUTURE PREDICTION, USING STOCHASTIC ADVERSARIAL BASED SAMPLING, FOR ROBOTIC CONTROL

Inventors:

JIANG LI (CN)
CHU CHAOQUN (CN)

Application Number:

PCT/CN2021/078444

Publication Date:

October 28, 2021

Filing Date:

March 01, 2021

Export Citation:

Click for automatic bibliography generation Help

Assignee:

UNIV SHANGHAI JIAOTONG (CN)

International Classes:

G06N3/08

Foreign References:

CN111507473A	2020-08-07
CN110097172A	2019-08-06
CN110874636A	2020-03-10

Other References:

LU, HAIWEI ET AL.: "Dynamic Network Structured Pruning via Feature Coefficients of Layer Fusion", PATTERN RECOGNITION AND ARTIFICIAL INTELLIGENCE, vol. 32, no. 11, 30 November 2019 (2019-11-30), pages 1051 - 1059, XP055859983, ISSN: 1003-6059
LI, ZHISHENG: "Design Space Exploration for Deep Learning Accelerator and Design and Implementation of the Accelerator in FPGA", CHINA MASTER’S THESES FULL-TEXT DATABASE, no. 02, 15 February 2020 (2020-02-15), pages 1 - 68, XP055860000

Attorney, Agent or Firm:

FANG & ASSOCIATES (CN)

Download PDF:

View/Download PDF PDF Help

Previous Patent: DISPLAY PANEL, DISPLAY DEVICE, AND METHOD FOR MANUFACTURING DISPLAY PANEL

Next Patent: PIXEL DRIVING CIRCUIT, DISPLAY SCREEN, AND TERMINAL