Bereiche | Tage | Auswahl | Suche | Downloads | Hilfe

DY: Dynamik und Statistische Physik

DY 15: Neuronale Netze

DY 15.1: Fachvortrag

Dienstag, 18. März 1997, 16:00–16:30, R1

Structure Formation in Neural Networks — •Hans-Otto Carmesin — Inst. f. Theoretische Physik, Universiät Bremen, 28334 Bremen

The brain’s 10¹⁶ synapses exhibit a more differentiated structure than genotype’s 10¹⁰ bits. What is the physics of synaptic structure formation? The incremental synaptic change is induced by subsequent firing of pre- & postsynaptic neuron. According to thermal membrane fluctuations, a Marcov-process in neuro-synaptic space is modeled. A field theory is obtained by averaging, adiabatic elimination of fast neuronal variables & integration in coupling space; so one gets a generalized free energy type expression & thus a solution in the sense of statistical physics & for non-equilibrium open systems. [1,2]      Successful theory applications include: formation of topology preservation, of orientation & direction preference & patterns - equivalent to Poisson-equation(!), ocular dominance patterns & eye movement coordination - even unsupervised learning induces sensor-motor coordination(!); learning of sequences & stimulus relations - hereby reaction times depend on spike structure & task difficulty(!); learning of counting without limitation; binding of stimuli to percepts; a schizophrenia model. Quantitative agreement is obtained between theory, empirical observations & simulations, new phenomena have successfully been predicted and self-organizing optical computers are proposed [3].  Lit.  [1] H.-O. Carmesin: Theorie neuronaler Adaption (Köster, Berlin 1994, ISBN 3-89574-020-9, 2nd Ed. 1996). [2] H.-O. Carmesin: Neuronal Adaptation Theory (Peter Lang, Frankfurt a.M. 1996). [3] H.-O. Carmesin: Self-organizing holographically realizable neural networks. ZKW-report10/96, Univ. Bremen 1996.