Mathematical Theory of Neural Networks

Abstract : This report focuses on fundamental theoretical issues relevant to the capabilities, performance, and limitations of artificial neural networks. For static (feedforward) networks, subjects of investigation included the study of error surfaces for least squares fitting, VC and other learning dimensions, representability questions, and function approximation. For dynamic (recurrent) nets, covered are questions dealing with parameter identification and modeling, realizability and other systems-theoretic issues, theoretical computational capabilities, and learning-theoretic issues.