Dynamics of interacting neural networks.pdfVIP

Dynamics of Interacting Neural Networks W Kinzel and R Metzler Institut f ur Theoretische Physik Universtitat Wurzbur g Am Hubland D Wurzbur g Germany I Kanter Minerva Center and Department of Physics Bar Ilan University Ramat Gan Israel July Abstract Complex bit sequences generated by a p erceptron that learns the opp osite of its own prediction are studied and the success of a student p erceptron trained on this sequence is calculated A system of interacting p erceptrons with a di rected ow of information is solved analytically A symmetry breaking phase transition is found with increasing learning rate A system of interacting p er ceptrons can develop a go o d strategy for the problem of adaptive comp etition known as the minority game Simple mo dels of neural networks describ e a wide variety of phenomena in neurobiology and information theory Neural networks are systems of elements interacting by adaptive couplings which are trained by a set of examples After training they function as content addressable asso ciative memory as classiers or as prediction algorithms Using metho ds of statistical physics many of these phenomena have b een elucidated analytically for innitely large single neural networks Many phenomena in biology so cial science and computer science may b e mo delled by a system of interacting adaptive algorithms One of the simplest of such mo dels is a system of interacting p erceptrons However nothing is known ab out general prop erties of such systems In this Letter we present analytic solutions of a system of interacting p erceptrons Firstly we stu