前馈神经网的定性学习理论

A new qualitative learning algorithm for feedforword neural networks is presented. In which weights are divided into rule types describing properties of weights and rule strength. The rule type of weights can be easily trained by back-propagation of superior contradiction. The qualitative learning theory is of many advantages, which lacked in traditional BP algorithm.The research provides necessary theory basis for the discovery of qualitative learning principle and the improvement of learning speed of neural networks. Some important researches on multi-mediea are also developed.

本项目提出和研究一种新颖的前馈神经网的定性学习算法，将权值用规则类型和规则强度表示，规则类型描述权值的性质；利用优势矛盾的反向传播快速学习各层权的规则类型。该定性学习理论具有传统的BP算法缺乏的许多优点如高速度和自适应能力。该研究为揭示突触类型的定性计算原理和提高神经网的学习速度提供必要的理论基础。.