Nonlinear factorization in sparsely encoded Hopfield-like neural networks
Anton Sirota, Alexander Frolov, Dušan Húsek · The European Symposium on Artificial Neural Networks · 1999
Moscow Institute of Physics and Technology*Institute of Higher Nervous Activity and Neurophysiology of the RAS**Institute of Computer Science AS of the CRThe problem of binary factorization of complex patterns in recurrent Hopfield-like neural network was studied both theoretically and by means of computersimulation. The number and sparseness of factors mixed in patterns cruciallydetermines the ability of an autoassociator to perform a factorization. Basing onexperimental data on memory and learning one may suggest, that there exists aneural system of intermediate storage of information, which fulfills the functionof binary factorization of the incoming polysensory information for its furthereffective storage in the form of elementary associatively bound factors. Wesuppose that field CA3 of the hippocampus possessing all properties of theautoassociative memory performs such function. This functional idea could befruitfully applied to various memory related tasks (e.g. spatial navigation) andlead to some critical experiments.