Reviews: Extracting low-dimensional dynamics from multiple large-scale neural population recordings by learning to predict correlations

The authors suggest a method to estimate latent low dimensional dynamics from sequential partial observations of a population. Using estimates of lagged covariances, the method is able to reconstruct unobserved covariances quite well. The extra constraints introduced by looking at several lags allow for very small overlaps between the observed subsets. This is an important work, as many imaging techniques have an inherent tradeoff between their sampling rate and the size of the population. A recent relevant work considered this from the perspective of inferring connectivity in a recurrent neural network (parameters of a specific nonlinear model with observations of the entire subset) [1].