Iterative Retrieval In Associative Memories By Threshold Control Of Different Neural Models

We investigate the retrieval properties of Hebbian auto-associative memories in the limit of sparse coding. Appropriately chosen threshold control strategies in nite size associa-tive memories increase the completion capacity for iterative retrieval (and even for the very fast two-step retrieval) above the asymptotic capacity for extremely large networks. We relate these results to a biologically motivated network consisting of excitatorily coupled cells which are controlled by a globally acting inhibitory interneuron. Choosing a homogenous coupling matrix and diierent excitatory single neuron types, we nd in an explicit numerical comparison, that the global behavior of spiking neurons in general is diierent from rate-function and probabilistic binary units. We also show that a network of spiking neurons with Hebbian coupling matrix is able to complete and segregate several distorted patterns that are simultaneously presented at the input. In this case the global dynamics falls into rhythmic activity and processes one input pattern per period, a behavior that might be related to rhythmic cortical activity as found, for example, in the visual cortices of cats and monkeys.