AUTONOMOUS DEVELOPMENT OF DECORRELATION FILTERS IN NEURAL NETWORKS WITH RECURRENT INHIBITION

We perform a quantitative analysis of information processing in a simp le neural network model with recurrent inhibition. We postulate that b oth excitatory and inhibitory synapses continually adapt according to the following Hebbian-type rules: for excitatory synapses correlated p re- and post-synaptic activity induces enhanced excitation; for inhibi tory synapses it induces enhanced inhibition. Following synaptic equil ibration in unsupervised learning processes, the model is found to per form a novel type of principal-component analysis which involves filte ring and decorrelation. In the light of these results we discuss the p ossible role of the granule-/Golgi-cell subnetwork in the cerebellum.