Carsen Stringer, Ph.D.
High-dimensional geometry of visual cortex
A neuronal population encodes information most efficiently when its stimulus responses are high-dimensional and uncorrelated, and most robustly when they are lower-dimensional and correlated. To determine the dimensionality of neural populations, we developed new computational techniques which allowed us to extract tens of thousands of neurons from calcium imaging data, and to estimate eigenvalues in the presence of neural variability. We found that the sensory-evoked population activity was high-dimensional, and correlations obeyed an unexpected power law: the nth principal component variance scaled as 1/n. We proved mathematically that if the eigenvalue spectrum was to decay more slowly then the population code could not be smooth, allowing small changes in input to dominate population activity. The theory also predicted larger power-law exponents for lower-dimensional stimulus ensembles, which we validated experimentally. This high-dimensional neural structure could not be easily captured by existing data visualization methods. We therefore developed a new visualization algorithm called Rastermap that explicitly enforces power-law scaling in the embedding. This mapping preserved high-dimensional, nonlinear relationships between neurons, providing useful visualizations for neuroscientists.
Carsen Stringer received her B.S. in Applied Mathematics & Physics at University of Pittsburgh. She completed her Ph.D. at the Gatsby Computational Neuroscience Unit at University College London, where she worked on neuronal network analysis and large-scale neural data analysis under the supervision of Kenneth D. Harris. Carsen is now a postdoctoral fellow at HHMI Janelia Research Campus under the supervision of Marius Pachitariu and Karel Svoboda working on mouse visual processing and biological image analysis.