No label defined (Q4382647)

This volume contains twenty collected articles on various mathematical aspects of neural networks. After an overview of computational, dynamical, and statistical perspectives on neural network theory each of these three topics is discussed in depth. The first one, computational perspectives on neural networks, deals with the processing problem in neural networks, and particular emphasis is laid on the parallelism inherent to neural networks. The approaches authors take range from discrete and analog computability theory over computational complexity theory to stochastic processes. Typical questions being discussed are: How do functions that can be computed by neural networks compare with those computed by Turing machines or other serial automata, or other networks of simple automata?; How are resource requirements affected by local parallelism, or by having threshold, as opposed to Boolean, operations at each unit?; What resources are required to determine whether weights exist for which a given architecture can reproduce a given finite set of data?; Can stochastic activation dynamics allow nets to globally optimize well-formedness of the activation state? What functions can be computed by networks of analog devices, and how do they compare with the functions computable by Turing machines? The second part of the book deals with dynamical perspectives on neural networks, i.e, the temporal or dynamic properties of activity (and, of weights) in networks that evolve over an extended period of time. These issues are discussed from the viewpoint of dynamical systems theory, statistical mechanics, function approximation theory, control theory, and time series analysis (dynamical systems theory, information theory). Typical questions studied are: What is the behavior (activation as well as learning dynamics) of neural networks?; Can global properties of (large) neural networks be understood by describing the neural network at a macrolevel in terms of variables that average over units, connections, and training patterns?; Can neural networks be effectively used as function approximators in the control of (nonlinear) dynamical systems?; Can neural networks succeed linear models as the method of choice for predicting time series values and for understanding their associated dynamical systems? In the third part, interest shifts to statistical perspectives on neural networks by focussing, in particular, on the learning problem. The approaches being considered include regularization theory, probability theory, information theory, hidden Markov modeling, computational learning theory and mathematical statistics. Typical questions under scrutiny are: Are neural networks suited to solving ill-posed problems (e.g., in inductive learning)?; Can Bayesian probabilistic analysis tell us the right error function activation dynamics for neural networks?; How can the outputs of complex systems be predicted or recognized?; How can such prediction/recognition be learnind?; Which classes of functions can be learned with resources growing only as a polynomial function of the size of the input?; How large a sample of training data is required to ensure that a network of given size will generalize well to unseen data?