Hardware-friendly STDP learning method and system based on threshold adaptive neurons

Abstract

The invention belongs to the technical field of spiking neural networks, and relates to a hardware-friendly STDP learning method and system based on threshold adaptive neurons, and the method comprises the following steps: S1, coding an input picture into a pulse sequence through frequency coding, and inputting the pulse sequence into a spiking neural network SNN; s2, receiving an input pulse by an excitation layer neuron of the SNN, accumulating a membrane potential, and when the membrane voltage reaches a threshold value, issuing the pulse and resetting the membrane potential; s3, the SNN inhibition layers are in one-to-one connection with excitation layer neurons, receive output pulses of the excitation layer neurons and inhibit the excitation layer neurons; s4, updating the excitatory synapse weight between the input layer and the excitation layer by adopting a hardware-friendly weight normalization method according to an STDP learning rule; and S5, after learning is completed, image recognition is carried out by using the pulse output sequence of the excitation layer neurons. According to the method, the resource consumption of the algorithm in hardware implementation can be reduced while the accuracy and the stability are ensured.

Description

technical field [0001] The invention belongs to the technical field of impulse neural networks, and relates to a hardware-friendly STDP learning method and system based on threshold self-adaptive neurons. Background technique [0002] Spiking Neural Network (SNN) is called the third-generation artificial neural network, in which spiking neurons transmit and calculate the spike sequence as effective information, and have unique characteristics in neuron model, synapse model, and pulse firing mechanism. Strong biological rationality, highly close to the real biological neural network. Hebb learning rules indicate that the connection strength of synapses changes with the changes of pre-synaptic and post-synaptic neuron activities, and neuroscience research has found that changes in synaptic weights are closely related to the precise timing of neurons firing pulses. The relative time difference of has a key effect on the change of weight direction and size. This learning rule ...

AI Technical Summary

Problems solved by technology

[0003] The standard weight normalization method of STDP contains division operations, which consumes a lot of resources during hardware implementation, and it is difficult to efficiently deploy on dedicated hardware devices with limited resources.

Images