Mathematics Colloquia and Seminars

Return to Colloquia & Seminar listing

Short-term memory is thought to be stored in the brain in the pattern of neural activity that persists following a transient stimulus. The memory of the stimulus is retained for as long as the pattern persists. For a network of neurons to be able to store memory of different stimuli, the network must be capable of maintaining a number of distinct patterns of persistent activity. This requires appropriate tuning of strengths of recurrent interactions between neurons in the network. The precise form of such interactions, however, is not well understood. Previous models of short-term memory networks assumed artificial constraints on both the patterns of connections and the nature of responses in the elements of the network. We overcome this limitation by developing a modeling framework where readily obtainable single-cell data (including stochastic noise) is used to predict the network connectivity. We apply this procedure to a network, known as a neural integrator, which calculates and stores the eye position resulting from a given sequence of eye velocity commands. We determine possible network architectures that reproduce the experimental performance of an intact network and of a partially inactivated network. We find that the best fitting network connectivities may be either recurrent or feedforward, but are nearly always sparse, even when the maximal connection strength is constrained. We show why sparseness may generally be expected in biologically realistic short-term memory networks. The present modeling framework is applicable in a wide variety of short-term memory settings.