Autonomous development of decorrelation filters in neural networks with recurrent inhibition

We perform a quantitative analysis of information processing in a simple neural network model with recurrent inhibition. We postulate that both excitatory and inhibitory synapses continually adapt according to the following Hebbian-type rules: for excitatory synapses correlated pre- and post-synaptic activity induces enhanced excitation; for inhibitory synapses it induces enhanced inhibition. Following synaptic equilibration in unsupervised learning processes, the model is found to perform a novel type of principal-component analysis which involves filtering and decorrelation. In the light of these results we discuss the possible role of the granule-/Golgi-cell subnetwork in the cerebellum.