Dynamics and Bifurcation of Neural Networks A Chapter in Handbook of Neural Networks edited.pdfVIP

Dynamics and Bifurcation of Neural Networks A Chapter in Handb o ok of Neural Networks edited by M Arbib Xin Wang Edward K Blum Computer Science Department Department of Mathematics University of California University of Southern California Los Angeles CA Los Angeles CA Intro duction While a feedforward neural network implements a function from its input layer to its output layer a recurrent neural network RNN denes a dynamical system whose dynamics is utilized to carry out a computation task such as implementing asso ciative memories pattern recognition and opti mization The practical ob jective of studying RNNs is to design network dynamics for p erforming a given task by adjusting network parameters This involves two ma jor issues in dynamical system theory namely dynamics and bifurcation Dynamics is concerned with the asymptotic b ehavior of the networks which includes limit sets eg xed p oints p erio dic orbits and chaotic invariant sets and their asymptotic stability eg stable unstable and saddle while bifurcation is concerned with how the dynamics changes as parameters are varied In many applications of RNNs in informa tion pro cessing such as asso ciative memories pattern recognition and classication combinatorial optimization segmentation and binding of ob jects etc limit sets of a network are used to repre sent the computational ob jects namely memories temp oral patterns optimal solutions or visual ob jects etc asymptotic stability of these limit sets is considered as asso ciativity faulttolerance generalization and robustness and bifurcation of these asymptotic structures as parameters chang