Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
SIGNAL PROCESSING METHOD FOR NEURON IN SPIKING NEURAL NETWORK AND METHOD FOR TRAINING SAID NETWORK
Document Type and Number:
WIPO Patent Application WO/2023/284142
Kind Code:
A1
Abstract:
A signal processing method for a neuron in a spiking neural network, and a method for training said network. Unlike a single spike mechanism that is presently commonly used, same is designed as a multi-spike mechanism. The signal processing method for a neuron comprises: a reception step: at least one neuron receives at least one input spike train; an accumulation step: a membrane voltage is obtained on the basis of a weighted sum of the at least one input spike train; an activation step: once the membrane voltage exceeds a threshold, the amplitude of a spike fired by a neuron is determined on the basis of a ratio of the membrane voltage and the threshold. In order to solve the problems of a training algorithm being inefficient and time-consuming due to an ever-increasing configuration parameter scale, the present network training method achieves highly efficient training of a spiking neural network by means of a multi-spike mechanism, a periodic exponential function surrogate gradient, and addition and suppression of a neuron activity level as loss, low power consumption of neuromorphic hardware can be sustained, and precision and convergence speed are also improved.

Inventors:
SHEIK SADIQUE ULAMEEN (CH)
XING YANNAN (CN)
WEIDEL PHILIPP (CH)
BAUER FELIX CHRISTIAN (CH)
Application Number:
PCT/CN2021/123091
Publication Date:
January 19, 2023
Filing Date:
October 11, 2021
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
CHENGDU SYNSENSE TECH CO LTD (CN)
International Classes:
G06N3/063; G06N3/08
Foreign References:
CN113255905A2021-08-13
CN111639754A2020-09-08
CN112465134A2021-03-09
CN113033795A2021-06-25
CN108304913A2018-07-20
US20200019850A12020-01-16
Attorney, Agent or Firm:
PURPLEVINE INTELLECTUAL PROPERTY (SHENZHEN) CO., LTD. (CN)
Download PDF: