ELSC Seminar Series

Modern recordings let us track large neural populations during behavior, but interpretation still hinges on finding principles that generalize across animals, tasks, and models. I will argue for a shift from describing activity only as continuous trajectories to treating it as discrete geometry: a structured set of metastable states and allowed transitions that can be compared across conditions and linked to circuit mechanisms. First, I will present evidence from large-scale recordings in the visual thalamus–cortex–midbrain circuit showing that task engagement constrains population activity, and that changes in cortical representational structure are largely predictable from the dimensionality and variance structure of its inputs. Second, I will describe a theory in which single episodes are stored as high-dimensional “path” traces written by a simple plasticity rule, enabling later recall by retracing the path through state space. Together, these stories show how discrete geometric descriptions provide compact, testable links from input statistics and learning rules to population representations.