Recording technology now allows us to record populations of neurons in multiple brain areas simultaneously. Traditionally, such recordings have been analyzed by identifying direct interactions between pairs of neurons, which provides a limited view of how distributed activity patterns in one area interact with those in another. We sought instead to use dimensionality reduction approaches to identify a small number of latent variables which summarize population activity patterns, and then to study the interaction of these latent variables. Specifically, we applied probabilistic canonical correlation analysis (pCCA) to populations of neurons recorded simultaneously in visual areas V1 and V2 of anesthetized macaque monkeys, while the animals were shown sinusoidal gratings of different orientations. We used pCCA to identify response subspaces of high correlation between the populations, termed "communication subspaces".

pCCA revealed that stimulus presentation produced a robust decorrelation of V1-V2 activity, compared to the spontaneous state. To understand the basis of this effect, we examined the relationship between trial-to-trial fluctuations in V1 activity and the V1-V2 communication subspace. We found that the average angle between V1 activity and the communication subspace increases during stimulus presentation, suggesting that V1 variability is 'filtered out' when V2 reads out V1 activity.

To account for possible delays in the communication between the two areas, we applied pCCA to time-shifted data, obtaining a multivariate cross-correlogram (mCCG) which shows the maximum correlation between the populations as a function of the time delay between them. Shortly after stimulus onset the mCCG included a clear feedforward component, with correlation being maximal for V2 responses that followed those in V1 by 2ms. Later in the evoked response, the mCCG revealed stronger correlation for V2 activity preceding V1. This suggests that the initial feedforward interaction is replaced later in the trial by a feedback interaction.

Our results suggest that representing distributed activity patterns with a small number of latent variables can be revealing for understanding the interactions between distinct neuronal populations. In particular, our analysis revealed novel dynamic interactions between V1 and V2, and suggests a sophisticated communication subspace which may mitigate the detrimental effects of V1 variability on V2.