Apparently high-dimensional spontaneous neural activity is locally low-dimensional in time

It has been commonly reported across species and brain regions that neural population dynamics are low-dimensional, leading to a broad range of theories on the significance of this structure [1, 2]. However, several recent studies have reported that spontaneous activity in rodent sensory regions is high-dimensional [3], presenting a challenge to explanations which posit low-dimensional dynamics as integral to neural computations. Here we hypothesized that the observation of high-dimensional population dynamics in these datasets results from neural activity progressing through a series of low-dimensional manifolds. That is, dynamics which seem to explore a high-dimensional subspace over a long period of time may be better described as dynamics which explore many low-dimensional subspaces over many short periods of time, where the high-dimensional space is composed of the union of the low-dimensional subspaces. We developed a novel statistical model and algorithm termed SPLAT to quantify the time-local dimensionality of neural population activity, defined as the number of active latent neural patterns during a small segment of time. We first established in synthetic data how SPLAT correctly recovers temporally sparse low-dimensional structure that appears high-dimensional when analyzed by PCA. We then applied this method to mouse V1 neurophysiology recordings from the Allen Brain Visual Behavior Neuropixels dataset [4], finding that SPLAT infers a small number of meaningful latent variables whose dynamics directly capture the temporally sparse stimulus structure of intermittent flashes of light in between periods of darkness. Finally, we found that these population recordings show evidence of local low dimensionality, with PCA estimating that 32 dimensions capture 85% of the variance, while SPLAT infers no more than 6 patterns are active during any given 10 ms bin. Together, these results suggest that apparently high- dimensional activity may represent passage through a sequence of low-dimensional neural computations.