Working memory: Patterns, variability and computational modeling

Current models of working memory assume that cortical working memory is achieved through sustained, reliable, stable increases in firing frequency in specific cue-excited subpopulations of cells. The aim of this thesis is to examine identify other changes that occur in cortical cell firing that occur during working memory, to characterize the variability in neuronal activity during working memory tasks, and to produce computational models incorporating alternative mechanisms that produce results that are more consistent with the physiological data.;We devise and implement a novel pattern detection algorithm to spike-trains recorded from primates performing a haptic delayed match-to-sample task. In general, the degree of patterning significantly increases during active memory. This increase in patterned firing is primarily due to an increase in patterning at high frequencies. Surrogate analysis suggests that the observed patterns may not be simple linear stochastic functions of firing frequency, and analysis of error trials suggests that these pattern changes are associated with behavior.;We explore the statistics and variability of neuronal activity during performance of working memory tasks. We find that the memory period signal and the memory-period frequency changes in cortical cells are smaller than generally considered. We also find that the behavior of individual cells varies substantially from trial-to-trial, and that firing frequency varies markedly over the course of a single trial. Furthermore, we show that delay-deactivation is common and similar in magnitude to delay-activation. These results are inconsistent with current models of working memory.;In most network models of working memory, network structure is assumed to be static over the course of a working memory trial. In contrast, physiological data raises the possibility that the synaptic plasticity may be induced during working memory tasks. We investigate the effects of implementing a Hebbian-type manipulation into a simple network model of working memory. We find that networks with Hebbian synaptic potentiation can produce persistent activation at firing rates and parameter variability consistent with experimentally observed cortical data. We also examine the effects of incorporating synaptic augmentation or Hebbian plasticity in a fully dynamic model of working memory. While augmentation decreases the amount of prior structure required for stable memory activity, it does not decrease the minimum frequency at which that activity occurs. In contrast, Hebbian plasticity permits memory signal to be stably maintained with realistic frequency changes.