Dynamic synapse neural network (DSNN): A new configuration based on wavelet filter bank and genetic algorithm
Hassan H. Namarvar, Jim‐Shih Liaw, Theodore W. Berger · The Journal of the Acoustical Society of America · 2001
A new DSNN architecture has been developed by the wavelet filter bank and the genetic algorithm (GA) training algorithm. The original DSNN [J.-S. Liaw and T. W. Berger, Hippocampus 6, 591–600 (1996)] was based on the integrate-and-fire-based neurons using Hebbian learning. Implementing input neurons of DSNN as wavelet filters has been shown to increase the retention of input information and the system performance in comparison to the integrate-and-fire neural network with a Hebbian learning algorithm. This study has shown improvements in achieving convergence of the neural network for difficult discrimination conditions. The Hebbian and anti-Hebbian learning rules do not take into account all of the system parameters which may have a significant impact on system performance. Hence we used the GA as the training algorithm to provide an efficient way for searching parameter spaces. This novel network has been tested by raw speech waveforms for a speech recognition task. The system performance during training phase was highly improved in comparison with the prior version of DSNN. [Work supported by DARPA.]