大规模神经网络的混合结构和立体表示系统—HYBNOTATION

HYBNOTATION A Hybrid Spatial Representation System and the Hybrid Organization of Large Scale Neural Networks

在文献［２４］中，我们提出了一个通用自适应神经网络框架ＧＡＦ，它具有许多优越性。但是，在利用神经网络模型构造具有特定功能的神经网络时，采用什么样的组织结构是很重要的。它决定着系统的代价和性能比。本文首先讨论了神经网络的自顶向下、自底向上和交互式三种基本结构及其局限性。然后提出了大规模神经网络的一个混合立体表示系统。接着通过两个实例，神经系统的反射弧和躯体感觉系统的定义，来说明上述混合立体表示系统对于神经网络微观及宏观结构描述的优越性以及如何用它来定义大规模神经网络的混合立体结构。

Paper[19],a general adaptive framework for neural networks hans been proposed.It has many advantages.But how to build a neural network with specific functions using the framework,and what organizational structures used are very important,because they dominate the cost and functions of the network.In this paper,three basic structures of neural networks. top-down, bottom-up and interactive and their limitations are discussed. Then,a hybrid spatial rep-resentation system for different scale of neural networks is presented.Two examples,the defining of the reflex are and the human somatic sensory system have been used to exemplfy how to use the hy-brid spatial representation system to define the micro-and macro-structures of large scale neural networks with hierarchical,hybrid (hierarchical and distributed) or pure parallel structures. The cvaluation of the hybrid representation system is also given.