The Role of Functional Modularity in the Neural Code Underlying Contextual Behavior

Human behavior is fundamentally contextual. For instance, the same person will put produce onto the shelves of a grocery store if they are at work, while they take produce off the shelves if they are shopping for dinner. Impairment of contextual behaviors are hallmarks of numerous neurological disorders, including post-traumatic stress disorder, schizophrenia, and addictive behavior. Understanding how the brain performs contextual behavior in health may be crucial to understanding the root causes of such disorders. One hypothesis is that contextual behavior is supported by modular neural representations in relevant sensory and decision-making brain areas, where the coordinated activity of a population of neurons is confined to specific subspaces that correspond to each context. This highly structured form of neural representation has been associated with contextual behavior before. However, it is not always associated with contextual behavior; instead, contextual behavior is sometimes associated with unstructured but high-dimensional neural representations. I hypothesize that the history of task learning shapes whether unstructured or modular geometry emerges and that the learned geometry then shapes the learning of further contextual tasks. I propose to test this hypothesis through a combination of mathematical theory, artificial neural network modeling, and analysis of neural population recordings from non-human primates. In Aim 1, I will determine when and why modular relative to unstructured representations emerge to support contextual behavior by building a theory of this emergence using artificial neural networks, and then testing the predictions of this theory in neural population recordings from sensory and decision-making brain regions as animals learn a sequence of contextual behaviors. In Aim 2, I will determine how modular relative to unstructured representations support the learning of future contextual behaviors – again by building a theory with artificial neural networks to be tested in non-human primate data. In Aim 3, I will determine how the flexibility of contextual behavior can come at the cost of behavioral precision across many different contexts, using a combination of mathematical theory and further analysis of additional neural population recordings. Together, these aims will elucidate the functional principles underlying large-scale modular organization in neural systems. Further, they will provide insight into how behavioral history influences both the performance of existing and the learning of novel contextual behaviors. During the K99 phase of this research program, I will be located at the Zuckerman Mind Brain Behavior Institute at Columbia University. My primary mentor, Dr. Stefano Fusi, will provide me with specific training in computational and theoretical neuroscience. My co-mentor, Dr. Roozbeh Kiani, will provide me with additional training in systems neuroscience and relevant analysis techniques. Their mentorship through this project and the extensive career development resources at the Zuckerman Institute will allow me to complete my training and begin my independent faculty career as a theoretical neuroscientist at a major research university.