# EDIT: it seems the post below me is actually correct, and my post is incorrect in this case.

Just as an aside here: "Blondeness" and "beard" are probably just labels the authors found correspond the most to the latent variables in this case. This means that there won't be a perfect translation between those words and what these variables directly respond to in the network.

So although the training data may have been biased with more smiling blonde people, it doesn't necessarily have to have been so. It might be that what this latent variable encodes just does something else in edge cases where there are few examples.