A Study on the Content, Format, and Implementation of Neural Representations That Underlie Flexible Human Cognition

Citation

Courellis, Hristos Spiridonos (2025) A Study on the Content, Format, and Implementation of Neural Representations That Underlie Flexible Human Cognition. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/edk1-kb22. https://resolver.caltech.edu/CaltechTHESIS:06122024-185146643

Abstract

Humans are the most capable cognitive generalists to walk the earth. They have a remarkable capacity for flexibility reallocating cognitive resources to rapidly acquire and execute an effectively infinite number of tasks. By utilizing the opportunity to record single-neuron activity in the frontal and temporal lobes of awake, behaving neurosurgical patients, we aim to elucidate the principles by which task representations are organized at the neural-circuit level to give rise to flexible cognition and behavior.

Our research program consists of four inter-related projects, each of which seeks to clarify the content, format, and single-neuron implementation of the representations that underlie different aspects of cognition and behavior that are uniquely human. In the first project, we demonstrate that the emergence of disentangled task representations in the hippocampus correlate with the ability of an individual to discover and perform inference on the state of latent context variables in their environment. In the second project, we describe differences in the temporal stability of instructed task representations in the hippocampus and medial frontal cortex, and show that they rely on persistent activity of single-neurons that lasts for 1-2 orders of magnitude longer than is typically studied in working-memory tasks. In the third project, we study the neural mechanisms of task-switching costs, and show that the state of medial frontal cortical context-representing neurons immediately following instructions is predictive of switching cost. In the fourth project, we evaluate the extent to which frontal cortical task representations inherit the compositional structure of natural language, and attempt to predict the neural representation of novel tasks as patients perform zero-shot generalization in a large task space.

Together, these projects constitute a first step in understanding the neural computations that underlie cognitive processing used by humans to solve complex, multi-task environments.

Thesis Files