A novel fast learning algorithms for time-delay neural networks

To counter the drawbacks that Waibel 's time-delay neural networks (TDW) take up long training time in phoneme recognition, the paper puts forward several improved fast learning methods of 1PW. Merging unsupervised Oja's rule and the similar error back propagation algorithm for initial training of 1PhW weights can effectively increase convergence speed, at the same time error firnction almost monotonyly descending. Improving error energy function. updating the changing of weights according to sue of output error, can increase training speed. From back propagation along layer, to average overlap part of back propagation error of first hidden layer along fiame, the training samples fiom a little to many gradually, convergence speed increases must faster. To multi-class phonemic modular 1PWs, we improve the architecture of Waible 's modular networks, and get an optimum modular lDhWs (OMWs) of tree structure to accelerate its learning, Its training time is less than Waibel's modular TDWs. The convergence speed increases about tens of times when the complexity of network increases just a little more, and the recognized rate is the same on the whole. Keywora3: Phoneme Recognition, Time-delay neural networks, Convergence speed, Optimum Module;