音素分类用函数链网的训练和结构优化

TRAINING AND STRUCTURAL OPTIMIZING OF FUNCTIONAL-LINK NEURAL NETWORK USED FOR PHONEME CLASSIFICATION

提出用于音素分类的函数链神经网的加速训练和结构优化算法,其基本思想是:在用改进BP法训练网络参数同时,逐步删剪对正确分类无用的联结和整个训练集中活度相关的节点,以获得最小化网络.实验表明,所提算法能显著改善BP训练性能,简化分类网结构,而很少影响其分类性能。

The training and structural optimizing algorithm of a functional-link neural network used for phoneme classification is proposed in this paper. The underlying ideas are: meanwhile the neural network is training by the improved BP algorithm, the links, which are useless for correct classification, and the nodes, their activities have most interreation over the whole training, are pruned for getting a network with the simplest structure. When it is used for training and optimizing the classification network of Chinese consonant ' b ', the input dimensions can be reduced from 90 to 18 while the correct recognition rate decreased from 96.98% to 95.29% only.