The global geometry of optimal neural representations
Year: 2026
Type: preprint
Abstract: Efficient coding hypothesizes that neural activity optimally represents information under resource constraints. The global structures of such optimal population codes are difficult to study theoretically, and previous works have usually resorted to either focusing on local properties like fine discrimination or placing restrictions on the parametric forms of neural tuning. We address the general problem by analytically deriving the neural representation that minimizes the all-pairs average classification error of a one-dimensional circular variable under various forms of neural response noise. Our method is based on formulating the space of neural response distributions as a statistical manifold whose geometry can be simplified through a generalized version of variance-stabilizing transformations. The theory fully characterizes the complete family of optimal population codes and allows us to make two novel predictions. First, an optimal code should allow the existence of oppositely peaked bimodal tuning curves. Second, there should be a correlation between spike count variability and tuning sharpness. We validate both of these predictions in the mouse head direction system.
Authors: Gengshuo John Tian, Brent Doiron
Institution: University of Chicago
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7236210128
