Date of Award
Doctor of Philosophy (PhD)
The brain’s remarkable computational properties arise from the collective activity of up to billions of densely interconnected neurons. Neurotechnologies to simultaneously measure the activity of many neurons have been steadily developed over the past decades, with mesoscopic optical imaging now reaching tens of thousands or more neurons in a single experiment. These data have only opened up additional fundamental questions on how the neuronal population code enables robust yet flexible computation from the highly variable activities of single neurons. Here, we utilized large-scale optical imaging to investigate how neuronal population dynamics are structured across diverse brain regions during spontaneous and variable behaviors. In the first part, we investigated the dimensionality and spatiotemporal structure of cortex-wide dynamics during spontaneous and uninstructed behaviors in the mouse, encompassing up to one million neurons recorded simultaneously and at multi-Hertz volume rates. While more than a decade of work has suggested that neuronal population dynamics appear to lie on low-dimensional manifolds that capture a large degree of neural variance and other sensorimotor features, recent evidence has suggested that ongoing dynamics in the brain exhibit higher dimensionality than previously appreciated. We found that the measured dimensionality of neuronal dynamics is even more high-dimensional than previously shown, and that the measured dimensionality scaled in an unbounded fashion with the number of recorded neurons. Within these dimensions, covarying ensembles of neurons were highly distributed across the entire dorsal cortex and relatively few were related to spontaneous behavior, suggesting that the majority of identified neural dimensions uniquely captured by large-scale recording were related to purely internal processing. Next, we switched our focus to the larval zebrafish in order to ask how highdimensional whole-brain dynamics produce population codes that are robust yet flexible enough to generate variable behaviors. To do so, we honed in on a regime of visual object size — between those that elicit hunting and avoidance behaviors — which induced maximum behavioral variability. We found that the visual encoding of object size is robust at the population level, despite the highly variable activity of single neurons. This robustness despite variability was due to the multi-dimensional geometry of the neuronal population dynamics: trial-to-trial “noise” modes were largely orthogonal to sensory encoding dimensions. Finally, we showed that these many of these noise modes were actually related to the larva’s behavior. Within this variability, we identified two brain-wide neuronal populations whose pre-motor activity predicted whether the larva would respond to a stimulus and, if so, which direction it would turn on a single-trial level. These populations were able to predict such single-trial behavior even seconds before the stimulus onset, suggesting they encoded time-varying internal biases that modulated the larva’s behavior, perhaps organizing behavior over longer timescales. In both the mouse and larval zebrafish, we found that neuronal population dynamics were extremely high-dimensional; mixed at the single neuron level, which can initially appear as “noisy” variability, but robust and highly structured at the population level; spatiotemporally structured in covarying ensembles that are distributed brain-wide; and dominated by the encoding of motor behavior and behaviorallyrelevant information, even in “non-motor” areas such as primary sensory regions.
Manley, Jason M., "On the High-Dimensional Geometry of Neuronal Population Dynamics" (2023). Student Theses and Dissertations. 734.
Available for download on Wednesday, April 30, 2025