Date of Award

Fall 10-1-2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Interdepartmental Neuroscience Program

First Advisor

Murray, John

Abstract

While humans can learn to perform many specific and highly specialized behaviors,perhaps what is most unique about human cognitive capabilities is their capacity to generalize, to share information across contexts and adapt to the myriad problems that can arise in complex environments. While it is possible to imagine agents who learn to deal with each challenge they experience separately, humans instead integrate new situations into the framework of the tasks they have experienced in their life, allowing them to reuse insight and strategies across them. Yet the precise forms of shared representations across tasks, as well as computational principles for how sharing of insight over learning multiple tasks may impact behavior, remain uncertain. The significant complexity in the problem of cognition capable of generalizing across tasks has been both an inspiration and a significant impediment to building useful and insightful models. The increasing utilization of artificial neural networks (ANN) as a model for cortical computation provides a potent opportunity to identify mechanisms and principles underlying multiple-task learning and performance in the brain. In this work we use ANNs in conjunction with human behavior to explore how a single agent may utilize information across multiple tasks to create high performing and general representations. First, we present a flexible framework to facilitate training recurrent neural networks (RNN), increasing the ease of training models on tasks of interest. Second, we explore how an ANN model can build shared representations to facilitate performance on a wide variety of delay task problems, as well as how such a joint representation can explain observed phenomena identified in the firing rates of prefrontal cortical neurons. Third, we analyze human multiple-task learning in two tasks and use ANNs to provide insight into how the structure of representations can give rise to the specific learning patterns and generalization strategies observed in humans. Overall, we provide computational insight into mechanisms of multiple-task learning and generalization as well as use those findings in conjunction with observed human behavior to constrain possible computational mechanisms employed in cortical circuits.

COinS