Thanks! Sure, I'll explain a bit. The method directly projects the inputs to a small number of latent variables, which are ordered in terms of explained covariance. So you'll know that the first one explains most of the covariance between data sets. Thus, the inputs 'connected' to it are most relevant to the relation between the two sets of data. On the other hand, the latter latent variables are of minor relevance and probably capture mostly noise (so random covariance). Here's where statistical testing of the identified latent variables comes in handy as a metric of what is signal and what is noise (and therefore should not be interpreted).
A nice side effect is that, because there are only a few latent variables, you can manipulate each of them individually and observe the effects on the inputs. In the case of facial expressions it's quite graphic.
I am a PhD student who researches the perception of emotional facial expressions. I am working on a statistical method which can utilize high-dimensional data for psychological research. Here's an app which illustrates this with face tracking data and emotion ratings: It lets you generate dynamic facial expressions at random or by setting emotion dimensions manually.