论文部分内容阅读
该文对音素识别中时延神经网络提出若干改进训练方法并与原方法进行实验对比研究,发现通过采取如下措施可有效地增加时延神经网络的收敛速率:(1)误差反传法初训权值.(2)从单极性输出改为双极性输出.(3)改变能量函数使权值修正根据输出误差的大小而改变.(4)将反传误差修正权值从按时延帧取平均改为按层进行.这些措施使收敛时间从原来的23小时另25分减少到45分钟,收敛速率提高数十倍之多,而网络复杂度增加很少.“,”The paper puts fond several improved methods of time-dely neural networks(TDNN) in phpnemerecognition. The comprison of proposed methods with early method shows that they are effective in increasingconvergence speed.of TDNN:(1) The crror baack propagation algorithm trains initially weights of the network.(2)The single-extreme output is replaced by the double-extreme output(3) Changing the energy function updatesweights according to output errors.(4) The weights update criterion of error back propagation was changed fromthe average weights of all corresponding time-delay frames to the layers. All these make the training time decrease from 23 hours and 25minutes to 45minutes .The convergence speed increases about tens of times when thecomplexity of network increases just a little more.