Abstract:
Hippocampal region dysfunction is suggested to have an important effect for thecognitive impairments observed in Alzheimer̕s disease. In some patients, hippocampus andnearby structures show atrophy while other brain structures appear intact. Hence, study of neural network models which can mimic biological and psychological findings is hoped tocontribute to our understanding of the underlying reasons and possible consequences ofhippocampal dysfunction. Therefore the main objective of this thesis work was to developan artificial neural network model that in many ways behaved like the hippocampal region. For this purpose we have used the cortico-hippocampal model of Gluck and Myers as thebasic model. The learning rule Gluck and Myers used in their original work wasbackpropagation. Hoping to get a more biologically plausible model, the learning rule waschanged to generalized recirculation (GeneRec). Furthermore, instead of using negative weights, the network was externally inhibited by two alternate methods: the kWTAinhibition and via additional inhibitory interneurons. Also, a weight bounding function wasapplied to the weight update rules.Addition of external inhibition and weight bounding functions to the network reduced the convergence characteristics of the network. Particularly cortico-cerebellar side of thenetwork could not converge with external inhibition. Therefore external inhibition wasabandoned for the cortico-cerebellar side. Although the hippocampal network couldconverge with kWTA, inhibition and weight bounding, rapid changes of activations of hippocampal network hidden layer neurons during training caused huge oscillations on thecortico-cerebellar output. Hence, external inhibition was abandoned also for thehippocampal network.The results of several representational differentiation and compression cases werefound comparable to the Gluck and Myers original work.|Keywords: Hippocampus, model, hippocampal atrophy, neural network, generalized recirculation
