摘要目的:提出一种新型脉冲神经元模型及其网络，描述其建模方法，并用计算机模拟验证其性能。方法:在充分考虑生物学适应性（激活电位阈值和不应期开关），及其对尖峰放电脉冲产生及其传导的动态调节机制基础上，在新型脉冲神经元模型中引入了突出后电位多通道滤波器，实现了输出电流及神经元突触强度的动态调节。提出基于自适应最小均方（LMS）的误差反向传播（BP）学习算法，并将其应用于尖峰放电神经网络的调节。结果:在自发噪声下，新型脉冲神经元模型的尖峰放电间期信号直方图满足泊松分布。通过2个新型脉冲神经元的简单连接，可以形成多种复杂的尖峰放电模式。新型脉冲神经元模型具有自发本征噪声的特征，能够形成复杂的周期尖峰放电模式。对于输入噪声控制，该模型的不应期与门限电位适应性参数的稳定性较好。刺激电流-尖峰放电脉冲频率间的线性关系较好。结论:所提出的新型脉冲神经元模型在自发噪声条件下能产生多种模式的振荡和相干振荡，这与生物神经元极其相似，能实现复杂的噪声信号处理。所采用的具有不同频带的多通道突触后电位滤波器，能使一些突触后电位信号变得平稳。所提出的基于于自适应LMS的BP学习算法克服了尖峰放电信号的瞬态变化特性导致的误差驱动学习算法无法应用的问题。

abstractsObjective: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.