Title:
ERROR CALIBRATION METHOD AND DEVICE FOR ANALOG NEURAL NETWORK PROCESSOR
Document Type and Number:
WIPO Patent Application WO/2019/232965
Kind Code:
A1
Abstract:
An error calibration method and device for an analog neural network (NN) processor. The method comprises: if algorithm update and/or error parameter adjustment are/is detected, parsing a network structure of an NN to obtain a trainable weight parameter of a full connection layer in the network structure (S101); using a stochastic gradient descent (SGD) algorithm to train the trainable weight parameter, wherein a loss value and a gradient in a learning process are subjected to logarithmic quantification, and the learning process is performed in a digital domain (S102); replacing back propagation in the learning process and multiplication operation used in the update of the trainable weight parameter with shift operation (S103); and storing a well-learnt weight parameter for the NN to calibrate an error of the processor according to the weight parameter (S104). The device executes the method. The energy and resource consumption of the analog NN processor can be reduced, so that the efficiency of the analog NN processor is improved.
Inventors:
JIA KAIGE (CN)
QIAO FEI (CN)
WEI QI (CN)
FAN ZICHEN (CN)
LIU XINJUN (CN)
YANG HUAZHONG (CN)
QIAO FEI (CN)
WEI QI (CN)
FAN ZICHEN (CN)
LIU XINJUN (CN)
YANG HUAZHONG (CN)
Application Number:
PCT/CN2018/104781
Publication Date:
December 12, 2019
Filing Date:
September 10, 2018
Export Citation:
Assignee:
UNIV TSINGHUA (CN)
International Classes:
G06N3/08
Domestic Patent References:
| WO2018034682A1 | 2018-02-22 |
Foreign References:
| CN107992842A | 2018-05-04 | |||
| CN107395211A | 2017-11-24 | |||
| CN107944458A | 2018-04-20 | |||
| CN105279554A | 2016-01-27 |
Attorney, Agent or Firm:
CN-KNOWHOW INTELLECTUAL PROPERTY AGENT LIMITED (CN)
Download PDF: