This idea is hardly surprising for anyone familiar with neural networks. If you train a deep auto-encoder (optimized for a truthful representation) you get a d

This idea is hardly surprising for anyone familiar with neural networks. If you train a deep auto-encoder (optimized for a truthful representation) you get a different set of abstract feature detectors than when you train a deep net that maps the perceptual input to some behavioral code.