Date of Award
Doctor of Philosophy (PhD)
This dissertation is aimed at understanding the cortical networks that maintain working memory information. By leveraging patterns of information degradation in spatial working memory encoding we reveal new neural mechanisms that support working memory function and challenge existing models of working memory circuits.
First we examine how interference from previous memoranda influences memory of a currently remembered location. We find that memory for a currently remembered location is biased toward the previously memorized location. This interference is graded, not all-or-none. Interference is strongest when the previous and current targets are close and activate overlapping populations of neurons. Contrary to the attractive behavioral bias, the neural representation of a currently remembered location in the frontal eye fields appears to be biased away from the previous target location, not toward it. We reconcile this discrepancy by proposing a model in which receptive fields of memory cells converge toward memorized locations. This reallocation of neural resources at task-relevant parts of space reduces overall error in the memory network but introduces systematic behavioral biases toward prior memoranda.
We also find that attractive behavioral bias asymptotically increases as a function of the memory period length. Critically, the increase in bias depends only on the current trial’s memory period. That is, the effect of the previous target progressively increases in the current trial after that target’s memory has become irrelevant. We modeled this finding using a two-store model with a transient but unbiased visual sensory store and a sustained store with constant bias. Initially behavior is driven by the veridical visual sensory store and is therefore unbiased. As the visual sensory store decays in the current trial, behavioral responses are increasingly driven by the sustained but biased store, leading to an asymptotic increase of behavioral bias with increasing memory period length.
Finally, we look at how memory activity is encoded over long (15 second) memory periods. Memory cells tend to turn on early in the memory period and stay active for a fixed amount of time. Most memory cells shut off prior to the end of the memory period. Within each cell, offset times are repeatable from one trial to the next. Across cells, offset times are broadly distributed throughout the entire memory period. Once a cell shuts off, it remains off for the rest of the memory period. On the one hand, these findings challenge the leading model for working memory, the attractor network framework, which predicts a single homogenous time course from all cells. On the other hand, the findings also show that the patterns of activity seen in memory circuits are much more structured than the heterogeneous patterns suggested by the leading competitors to the attractor models. Our findings are not predicted by current models of working memory circuits and indicate that new network models need to be developed.
Chair and Committee
Lawrence H Snyder
Todd Braver, Vitaly Klyachko, Camillo Padoa-Schioppa, Richard Abrams,
Papadimitriou, Charalampos, "Neural Mechanisms of Working Memory Cortical Networks" (2015). Arts & Sciences Electronic Theses and Dissertations. 586.
Permanent URL: https://doi.org/10.7936/K7H41PM0