On the Storage Capacity of an Abstract Cortical Model with Silent Hypercolumns

In this report we investigate the storage capacity of an abstract generic attractor neural network model of the mammalian cortex. This model network has a diluted connection matrix and a fixed activity level that is independent of network size. We develop an analytical model of the storage capacity for this type of networks when they are trained with both the Willshaw and Hopfield learning-rules. Experimentally we investigate three different learning-rules, the two mentioned and the BCPNN. We propose a new method for coding arbitrarily sparse patterns into this network, which allows some of the hypercolumns to be silent, i.e. they do not send any information to other hypercolumns. We find that silent hypercolumns cannot be used together with the Hopfield learning-rule. We show that silent hypercolumns increases the storage capacity and flexibility of this type of networks.