This paper is mainly devoted to the principle and the implemental algorithms ofqualitative learning for feed-forward neural networks (FNNs). A general discussion forlearning methods of FNNs and an introduction to logic inference are firstly given. Then,according to the analysis of networks and the evolution of their weights, a learningprinciple, based on superior contradiction reversion, is presented, which demonstratesthe partition of the synaptic quality and the synaptic strength. And a new learningalgorithm for binary FNNs, which possesses the property of high speed, dynamicnetwork structure and the adaptable ability for increasing learning patterns, is proposed.Finally, a perspective of learning methods for multi-output binary FNNs and thequalitative learning for real-value networks are given.
|