Search Papers | Poster Sessions | All Posters
Poster C23 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Neural Manifold Capacity Captures Representation Geometry, Correlations, and Task-Efficiency Across Species and Behaviors
Chi-Ning Chou1 (), Luke Arend2, Albert Wakhloo1,3, Royoung Kim4, Will Slatton2, SueYeon Chung1,2; 1Flatiron Institute, 2New York University, 3Columbia University, 4Sungkyunkwan University
Relating the coordinated activity of neurons to cognitive functions is a fundamental challenge in neuroscience. While experimental evidence indicates these neuronal populations act as core computational units, quantifying these population codes in relation to their functional roles has remained elusive. Prior approaches based on representational geometries, functional alignment or dimensionality reduction have faced limitations in robustly linking neural population structure to computation across scales and modalities. Here, we fill this gap by introducing effective Geometric measures from Correlated Manifold Capacity theory (GCMC), a framework that employs analytical methods from statistical physics, to connect the geometry of neural population activities to readout performance, thereby quantifying coding efficiency. Applying this to diverse neural recordings across organisms and tasks, we demonstrate multi-scale analyses previously inaccessible. These include tracking changes in coding efficiency and geometry across brain regions, revealing task-relevant manifold dynamics over time, and characterizing representational changes during learning. The geometric measures serve as interpretable descriptors relating the structure of coordinated neural population activity to embedded computations. Our framework provides a general and principled approach for mapping neural population codes to their functional roles, enabling data-driven insights into the neural underpinnings of perception and behavior.
Keywords: Neural representations Coding efficiency Population geometry