音素分类用函数链网的训练

Training 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改进训练方法。其基本思想是：用正反例均衡、样本跳转、目标函数修正、学习率自动调整和样本渐增等改善常规的BP算法，以提高其分类性能和收敛速度。文中还提出了一种形象和有效地评价分类器性能的方法—正反例样本分布直方图。实验表明，所提算法能显著改善BP训练性能。将其与结构优化算法结合用于训练和优化汉语辅音‘b’的分类网，可将输入维数由90维压缩到18维，而对其训练集的正识率仅由96.98％减为95.29％。

An improved training algorithm is proposed for functional-link neural network used for phoneme classification.The popular BP algorithm is improved by introducing an equilibrium between the positive and negetive samples,the skipping of the samples,the renewal of the target function,the adaptive adjusting of the learning rate and teh sample-increasing training for improving its classification performance and convergence rate.A new method for visually and effectively evaluation the performance of a classifier - the histogram of the distribution of the positive and negative samples is also proposed in this paper.It is shown by experiment that this algorithm can improve the performance of the BP training remarkably,and reduce the input dimensions of the classification net of the chinese consonant'b' form 90 to 18 with the correcognifion rate decreased from 96.98% to 95.29% only when it is combined with a structural optimizing algorithm.