Abstract

Recurrent neuronal networks of excitatory and inhibitory neurons are deemed to contribute significantly to sensory processing. Yet, the exact nature of this contribution and the mechanisms underlying it remain elusive. By studying links between neuronal dynamics and functional properties of recurrent networks in network models, we aim to elucidate the mechanisms responsible for some fundamental computations in primary sensory cortices. Starting with randomly connected networks lacking any feature-specific structure, e.g. corresponding to the visual cortex of rodents at eye opening, we analyze the input-output transfer function of the network in response to tuned input. Once the initial selectivity is established in the network, accounting for synaptic plasticity of excitatory and inhibitory connections can in fact lead to the emergence of feature-specific connectivity. This is consistent with experimental reports on the sequence of events during the emergence of orientation selectivity and specific connectivity. Analyzing the neuronal dynamics of the networks before and after synaptic refinement reveals further functional benefits of synaptic specificity, including the emergence of reverberating activity and pattern completion. Random and specific networks may, therefore, be responsible for different aspects of sensory processing at different developmental stages, as the computational analysis of neuronal dynamics and synaptic plasticity in large-scale recurrent networks reveals.