"Short term memory and coding in stochastic neural networks"

Short term memory is hypothesized to reside in the ongoing activity of neurons in the brain. Some important theoretical questions, which are raised in this context, are: How can ensembles of neurons produce slow dynamics, relative to the typical time scales that characterize the dynamics of single neurons? How faithfully can the activity in such ensembles represent information about the past? I will discuss these questions mainly in the context of brain areas that encode memories of continuous variables, such as an orientation or a position. I will argue that there is a fundamental relationship between the ability of a stochastic neural network to maintain a stable representation of a continuous variable, and another theoretical question which is often studied in relation to neural noise: How accurately can a population of neurons encode a continuous variable - as quantified by the accuracy of readout by an external observer of the neural activity?