A GENERALIZED GOODNESS CRITERION FOR UNSUPERVISED NEURAL LEARNING OF VISUAL PERCEPTION

作     者：Liu Jianqin(College of Information Engineering, Central South University of Technology, Changsha 410083, China) 

作者机构：1. College of Information Engineering Central South University of Technology 410083 Changsha China 

出 版 物：《Journal of Central South University》 (中南大学学报（英文版）)

年 卷 期：1996年第8卷第2期

页      面：63-67页

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

主　　题：visual perception unsupervised learning neural network 

摘      要：Unsupervised learning plays an important role in the neural networks. Focusing on the unsupervised mechanism of neural networks, a novel generalized goodness criterion for the unsupervised neural learning of visual perception based on the martingale measure is proposed in the paper. The differential geometrical structure is used as the framework of the whole inference and spatial statistical description with adaptive attribute is embedded in the corresponding nonlinear functional space. Consequently the integration of optimization process and computational simulation with the NeoDarwinian paradigm is obtained. And the generalization of the guidance for the evolutionary learning in the neural net framework, the convergence of the goodness and process of the evolution guaranteed by the mathematical features are discussed. This criterion has generic significance in the field of machine vision and visual pattern classification.