In fact almost the opposite! Unsupervised pre-training learns the statistical invariants in the inputs and these become the features. This is in contrast to ha

In fact almost the opposite! Unsupervised pre-training learns the statistical invariants in the inputs and these become the features. This is in contrast to hand engineering features for your particular domain which, unsurprisingly, doesn't generalise well.