Neural network learns to identify criminals by their faces

Learned to distinguish criminals who got arrested and photographed from people who didn't.

I expect the network is observing a lack of facial symmetry.

It is fascinating to watch ML retread a hypothesis humans came up with.

What I mean is, those criminals who did not get caught and but did get photographed are happily smiling in the not-criminal training set.

So the label of "criminal" is over general.

Neural network learns to identify criminals by their faces (technologyreview.com)