Search Papers | Poster Sessions | All Posters
Poster C41 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Universality in mouse and human visual cortex: relating covariance to the spatial structure of latent dimensions
Raj Magesh Gauthaman1 (), Brice Ménard1, Michael F. Bonner1; 1Johns Hopkins University
Recent work has revealed the high-dimensional structure of visual cortex responses to natural images in both mice and humans, where stimulus-related variance is distributed as a power law over thousands of latent dimensions. Here, we characterize the covariance spectra of two datasets containing V1 responses to thousands of visual stimuli measured at two very different scales: mouse calcium imaging and human fMRI. We find that the power-law exponent α characterizing the spectral decay varies substantially across experiments, contradicting previous claims of universality and optimality in the power law exponents of visual cortex. However, we also discover a striking pattern where variance along a latent dimension is directly related to its spatial scale — a measure of how strongly neighboring neurons co-activate. When viewed through this lens, the spectra of the mouse and human neural activations show striking similarities, suggesting that both visual systems represent natural images in similar ways. Our results demonstrate that analyzing the spatial scale of latent modes of variation might be a more fundamental way to quantify the covariance structure of neural representations.
Keywords: visual cortex dimensionality human fMRI mouse calcium imaging