Low-dimensional representations can capture structure in neural dynamics data, but it is unclear whether additional structure is being missed, especially when larger populations are sampled. Manley et al. have imaged the activity of up to a million neurons in dorsal cortex of awake head-fixed mice to directly measure how neural dimensionality scales with population size. Using shared variance component analysis, they found unbounded scaling of neural dimensionality with increasing numbers of sampled neurons, suggesting that there may be very high-dimensional signals in the neural dynamics. Behavior-related activity was encoded in a low-dimensional space mostly within the first 16 components, which accounted for only about half of the total neural variance, with hundreds of neural dimensions remaining unlinked to behavior or sensory variables. The higher-dimensional dynamics displayed a continuum of timescales and spatial patterns of distribution across cortex, and further work is needed to uncover the nature of these signals.

Original reference: Neuron https://doi.org/10.1016/j.neuron.2024.02.011 (2024)