Signal processing method and apparatus based on spiking neural network

Patent metadata

Field	Value
Publication number	US-11308386-B2
Application number	US-201916423750-A
Country	US
Kind code	B2
Filing date	May 28, 2019
Priority date	Nov 28, 2016
Publication date	Apr 19, 2022
Grant date	Apr 19, 2022

Abstract

A signal processing method and apparatus includes determining a first signal F 1 (t) output by a first neuron, processing the first signal F 1 (t) using q orders of synapse weight parameters w q (t), w q−1 (t), . . . , w 1 (t) to obtain a second signal F 2 (t), and inputting the second signal F 2 (t) to a second neuron, where the second neuron is in a layer immediately subsequent to that of the first neuron.

First claim

What is claimed is:

1. A signal processing method implemented in a spiking neural network and comprising:
receiving, by a memristor, a first signal F 1 (t) output by a first neuron;
processing, by the memristor, the first signal F 1 (t) using q orders of synapse weight parameters w q (t), w q−1 (t), . . . w 1 (t) to obtain a second signal F 2 (t), wherein a speed at which an initial function w x+1 0 (t) met by an (x+1) th -order synapse weight parameter of the q orders of synapse weight parameters attenuates with time t is higher than a speed at which an initial function w x 0 (t) met by an x th -order synapse weight parameter attenuates with the time t, wherein q is a positive integer greater than 1, and wherein 1≤x≤q−1; and
outputting, by the memristor, the second signal F 2 (t) to a second neuron, wherein the second neuron is a next-layer neuron of the first neuron,
wherein the first signal F 1 (t) comprises a first sub-signal F 1 (t 1 ) output by the first neuron at a moment t 1 ,
wherein a q th -order synapse weight parameter w q (t 1 ) at the moment t 1 meets the following condition:
w q ( t 1 )= w q ( t 1 −1)+Δ F q ( t 1 ),

wherein w q (t 1 −1) represents a q th -order synapse weight parameter at a moment t 1 −1, the moment t 1 −1 is a previous moment of the moment t 1 , and ΔF q (t 1 ) is an update quantity based on a learning rule and a stimulation effect at the moment t 1 of the first sub-signal F 1 (t 1 ) on the q th -order synapse weight parameter w q (t 1 −1) at the moment t 1 −1,
wherein an x th -order synapse weight parameter w q (t 1 ) at the moment t 1 meets the following condition:
w q ( t 1 )= w q ( t 1 −1)+Δ F x ( t 1 )+Δ w x,x+1 ( t 1 )+Δ w x,x+2 ( t 1 )+ . . . +Δ w x,q ( t 1 ),

wherein w x (t 1 −1) represents an x th -order synapse weight parameter at the moment t 1 −1 and ΔF x (t 1 ) is an update quantity based on the learning rule and the stimulation effect at the moment t 1 of the first sub-signal F 1 (t 1 ) on the x th -order synapse weight parameter w x (t 1 −1) at the moment t 1 −1, and
wherein Δw x,x+1 (t 1 ), Δw x,x+2 (t 1 ), . . . , Δw x,q (t 1 ) are respectively quantities of impact of an (x+1) th -order synapse weight parameter, an (x+2) th -order synapse weight parameter, . . . , and the q th -order synapse weight parameter that are at the moment t 1 on the x th -order synapse weight parameter w x (t 1 ).

2. The signal processing method of claim 1 , wherein an initial function w 1 0 (t) of a first-order synapse weight parameter of the q orders of synapse weight parameters does not attenuate with time.

3. The signal processing method of claim 1 , wherein an (x+i) th -order synapse weight parameter w x+1 (t 1 ) at the moment t 1 is greater than or equal to a threshold of an (x+i) th -order synapse weight, and wherein w x,x+i (t 1 ) is not 0.

4. The signal processing method of claim 1 , wherein an (x+i) th -order synapse weight parameter w x+i (t 1 ) at the moment t 1 is less than a threshold of an (x+i) th -order synapse weight, wherein Δw x,x+i (t 1 ) is equal to 0, and wherein i=1, 2, . . . , or q−x.

5. The signal processing method of claim 1 , wherein the learning rule is a learning rule based on a biological feature or a supervised learning rule based on an error back propagation mechanism.

6. The signal processing method of claim 1 , wherein the first signal F 1 (t) comprises a first sub-signal F 1 (t 1 ) output by the first neuron at a moment t 1 , and a q th -order synapse weight parameter w q (t 1 ) at the moment t 1 meets the following condition:
w q ( t 1 )= w q ( t 1 −1)+Δ F q ( t1 ),

wherein w q (t 1 −1) represents a q th -order synapse weight parameter at a moment t 1 −1, the moment t 1 −1 is a previous moment of the moment t 1 , and ΔF q (t 1 ) is an update quantity determined based on a learning rule and a stimulation effect at the moment t 1 of the first sub-signal F 1 (t 1 ) on the q th -order synapse weight parameter w q (t 1 −1) at the moment t 1 −1.

7. The signal processing method of claim 1 , wherein an (x+1) th -order synapse weight parameter w x+1 (t 1 ) at the moment t 1 is greater than or equal to a threshold of an (x+1) th -order synapse weight, wherein an x th -order synapse weight parameter w x (t 1 ) at the moment t 1 meets the following condition:
w x ( t 1 )= w x ( t 1 −1)+Δ w x,x+1 ( t 1 ),

wherein w x (t 1 −1) represents an x th -order synapse weight parameter at the moment t 1 −1, and wherein Δw x,x+1 (t 1 ) is a quantity of impact of the (x+1) th -order synapse weight parameter at the moment t 1 on the x th -order synapse weight parameter w x (t 1 ).

8. The signal processing method of claim 7 , wherein an (x+1) th -order synapse weight parameter w x+1 (t 1 ) at the moment t 1 is less than a threshold of an (x+1) th -order synapse weight, and wherein an x th -order synapse weight parameter w x (t 1 ) at the moment t 1 meets an initial function w x 0 (t 1 ).

9. The signal processing method of claim 1 , wherein processing, by the memristor, the first signal F 1 (t) using the q orders of synapse weight parameters w q (t), w q−1 (t), . . . w 1 (t) to obtain a second signal F 2 (t) comprises determining, by the memristor, a product of the first signal F 1 (t) and a first-order synapse weight parameter w 1 (t) as the second signal F 2 (t).

10. A spiking neural network comprising:
a first neuron configured to output a first signal F 1 (t);
a memristor coupled to the first neuron and configured to:
receive the first signal F 1 (t) output by the first neuron;
process the first signal F 1 (t) using q orders of synapse weight parameters w q (t), w q−1 (t), . . . w 1 (t) to obtain a second signal F 2 (t), wherein a speed at which an initial function w x+1 0 (t) met by an (x+1) th -order synapse weight parameter of the q orders of synapse weight parameters attenuates with time t is higher than a speed at which an initial function w x 0 (t) met by an x th -order synapse weight parameter attenuates with the time t, wherein q is a positive integer greater than 1, and wherein 1≤x≤q−1; and
output the second signal F 2 (t); and

a second neuron coupled to the memristor and configured to receive the second signal F 2 (t), wherein the second neuron is a neuron in a layer immediately subsequent to a layer of the first neuron,
wherein the first signal F 1 (t) comprises a first sub-signal F 1 (t 1 ) output by the first neuron at a moment t 1 ,
wherein a q th -order synapse weight parameter at the moment meets the following condition:
w q ( t 1 )= w q ( t 1 −1)+Δ F q ( t 1 ),

wherein w q (t 1 −1) represents a q th -order synapse weight parameter at a moment t 1 −1, the moment t 1 −1 is a previous moment of the moment t 1 , and ΔF q (t 1 ) is an update quantity based on a learning rule and a stimulation effect at the moment t 1 of the first sub-signal F 1 (t 1 ) on the q th -order synapse weight parameter w q (t 1 −1) at the moment t 1 −1,
wherein an x th -order synapse weight parameter w x (t 1 ) at the moment t 1 meets the following condition:
w x ( t 1 )= w x ( t 1 −1)+Δ F x ( t 1 )+Δ w x,x+1 ( t 1 )+Δ w x,x+2 ( t 1 )+ . . . +Δ w x,q ( t 1 ),

wherein w x (t 1 −1) represents an x th -order synapse weight parameter at the moment t 1 −1 and ΔF x (t 1 ) is an update quantity based on the learning rule and the stimulation effect at the moment t 1 of the first sub-signal F 1 (t 1 ) on the x th -order synapse weight parameter w x (t 1 −1) at the moment t 1 −1, and
wherein Δw x,x+1 (t 1 ), Δw x,x+2 (t 1 ), . . . , Δw x,q (t 1 ) are respectively quantities of impact of an (x+1) th -order synapse weight p

Assignees

• Huawei Tech Co Ltd

Inventors

• Wang Ming
• Yan Zheng
• Xue Xijun

Classifications

• G06N3/065

  Analogue means · CPC title

• G06N3/049 Primary

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N3/02

  Neural networks · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title