There's a pretty strong literature around retraining just the top couple layers of a deep net to Target a slightly different objective.

The interesting possiblity opened here is training new feature processing frontends to work with an established 'conceptual' backend.