Improving neural networks for pattern recognition and function approximation
Ximin Zhang · DR-NTU (Nanyang Technological University) · 1999
This thesis studies various issues related to artificial neural networks for pattern recognition and function approximation with the aim to enhance its capability and to improve its performance. It proposes a novel method for globally finding good minima and optimizing the Correct Classification Rate(CCR), and a novel algorithm for network construction and weight initialization. The thesis also an-alyzes the fundamentals of Time-Delay Neural Network(TDNN) and presents an augmented TDNN (ATDNN) for frequency and scale invariant sequence classification.