Learning temporal representations in cortical networks through reward dependent expression of synaptic plasticity
The neural basis of the brain's ability to represent time, which is an essential component of cognition, is unknown. Despite extensive behavioral and electrophysiological studies, a theoretical framework capable of describing the elementary neural mechanisms used by biological neural networks to learn temporal representations does not exist. It is commonly believed that the underlying cellular mechanisms reside in high order cortical regions and there is an ongoing debate about the neural structures required for temporal processing. Recent experimental studies report sustained neural activity that can represent the timing of expected reward in low-level primary sensory cortices, suggesting that temporal representation may form locally in sensory areas of the cortex. This thesis proposes a theoretical framework that explains how temporal representations of the type seen experimentally can be encoded in local cortical networks and how specific temporal instantiations can be learned through reward modulated synaptic plasticity. The proposed framework asserts that the mechanism responsible for encoding the observed temporal intervals is long-term synaptic potentiation between neurons in a recurrent network. Analytical and numerical techniques are used to demonstrate that the model is sufficient to allow näive networks of both linear and non-linear neurons to encode and reliably represent durations specified by external cues during a training period. Analysis of a non-linear spiking neuron model is accomplished using a mean-field approach. The form of temporal learning described has specific implications that can be confirmed experimentally and these predictions are highlighted. Experimental support for a central component of the model is presented and all of the the results are discussed in relation to current experimental and computational work.