Abstract:Objective:To study a new type of impulse neuron model and its network, describe its modeling method, and verify its performance by computer simulation.Methods:Based on full consideration of biological adaptability (activation potential threshold and refractory period switch) and its dynamic regulation mechanism for the generation and transmission of spike discharge pulses, a post-potential multi-channel filter was introduced in the new impulse neuron model. The filter can realize the dynamic adjustment of output current and neuron synaptic strength. An error back-propagation (BP) learning algorithm based on adaptive least mean square (LMS) was proposed, and applied to the regulation of spike discharge neural networks.Results:Under spontaneous noise, the pulsation interval signal histogram of the new impulse neuron model satisfies Poisson distribution. Through the simple connection of two new pulsed neurons, a variety of complex spike discharge patterns can be formed. The new impulse neuron model has the characteristics of spontaneous intrinsic noise, can form complex periodic spike discharge patterns. For input noise control, the refractory period and threshold potential adaptability parameters of the new impulse neuron model has good stability. The linear relationship between the stimulation current and the frequency of the spike discharge pulse is good.Conclusions:The proposed new model can generate multiple modes of oscillation and coherent oscillation under the condition of spontaneous noise, which is very similar to biological neurons and can realize complex noise processing. The multi-channel post-synaptic potential filters with different frequency bands can make some post-synaptic potential signals become smooth. The proposed BP learning algorithm based on adaptive LMS can overcome the problem that the error-driven learning algorithm cannot be applied due to the transient characteristics of the spike discharge signal.

Activity-dependent Neuron Model for Noise Resistance

Exploiting Noise As a Resource for Computation and Learning in Spiking Neural Networks

Noise is the Fatal Poison: A Noise-aware Network for Noisy Dataset Classification

Dynamic Behavior of Artificial Hodgkin-Huxley Neuron Model Subject to Additive Noise.

Noise Optimization in Artificial Neural Networks

Adaptive Regulation for Noise-Aided Weak Signal Detection in Spiking Neurons

Noise Helps Optimization Escape from Saddle Points in the Synaptic Plasticity

Enhance DNN Adversarial Robustness and Efficiency via Injecting Noise to Non-Essential Neurons

Noise Helps Optimization Escape from Saddle Points in the Neural Dynamics.

AGN: A Power-Adaptive and Growth-Capable Neuron Model and Circuit Implementation

Improving the Antinoise Ability of DNNs via a Bio-Inspired Noise Adaptive Activation Function Rand Softplus

Understanding and mitigating noise in trained deep neural networks

Turn Down that Noise: Synaptic Encoding of Afferent SNR in a Single Spiking Neuron

How Neuronal Noises Influence the Spiking Neural Networks’s Cognitive Learning Process: A Preliminary Study

Noise and signal: A model-based study of stochastic effects in neural systems

Alternating Excitation-Inhibition Dendritic Computing for Classification

Noise Adaptor in Spiking Neural Networks

Using noise to probe recurrent neural network structure and prune synapses

Research on a new impulse neuron model and its neural network

Aadaptive Signal De-Noising Based on Feedback Networks and Counterpropagation Network

Novel Classification Algorithms Inspired by Firing Rate Stochastic Resonance