基于生物的神经网络的理论框架──神经元模型

The Biology-Based Theoretical Frame For the NeuralNetwork-Neuron Model

由于人工神经网络（ＡＮＮ）的种类繁多、行为各异，因此有必要建立一个统一的建立在生物基础上的理论框架。以便进一步发展ＡＮＮ的各种高级应用。这个框架由神经元模型、突触的可塑性以及它的学习算法、神经网络的结构三部分组成。本文是它的第一部分，神经元模型是由建筑在神经生理基础之上的九个特性组成，可以定量描述包括：（１）空间总和效应；（２）时间总和效应；（３）阈值特性；（４）不应期；（５）适应性；（６）兴奋与抑制特性；（７）延时；（８）输出特性；（９）传导特性。这些特性是从长期的生理实验中获得的，也可以由Ｈ－Ｈ方程经计算机模拟获得［１］，我们作了这些模拟。由这九个特性构成的神经元模型不便于作统一的数学分析。其中某些特性在某些情况下是不重要的；如传导特性，适应性等。而另一些特性如空间总和效应，输出特性是较为重要的。衡量目前流行的各种神经网络模型，发现都没有超出九个特性［２］。应用这些特性于实践及各种仿真研究中，都获得良好结果如联想记忆及ＥＣＧ的数据压缩等。

In order to manifest the richness and variety of artificial neural networks (ANN), their intelligent behavior, and to develop advanced applications of ANN. it is necessary to build theoretical frame for the ANN Here presented is the first part of the fiame, which includes the following three parts: neuron model, plasticity of synapses and their leaening algorithms, and the architecture of the neural network. The neuron model is composed of nine features based on neurophysiology, including: ( 1 ) spatial summation effect (2) tempofal summation effect (3) the characteristics of the threshold; (4) non-respond duration; (5) adaptability; (6) excitation and characteristics; (7) delay: (8) characteristics of the output. (9) characteristics of conductivity. These features are obtained by long-term physiological experiments, or by H-H equation with computer simuiation which we carried out The neuron models, which have these nine neuron properties, are not convenient for mathematical analysis. In some cases. some characters of them are not important,such as conductivity, adaptability, etc. while some other characters (spatal summation effect, output character, etc.) are more important. If we review the popular ANN. models with the nine features, we can find that all of them are within these features. We used these features in practice and simulation research, satisfactory results were obtained, such as associated memory and data compression in ECG.