If we assume that humans use different brain structures to learn/reason symbolically and sub-symbollically, then I think it is natural to assume the following:

• Examples are better suited to train a brain part that deals with sub-symbol representations of the world.
• Existing symbols are better suited to operate on by a brain part that deals with symbols.

And it seems plausible that there are persons with brains better suited the the former and other to the latter. So there is no conflict but a 'it depends'.

I couldn't quickly find a good reference for symbol vs. sub-symbol learning in humans. You have to take this from AI: http://en.wikipedia.org/wiki/Artificial_intelligence#Sub-symbolic

A long time ago I wrote something about this distinction which explains why and when examples help here: http://c2.com/cgi/wiki?FuzzyAndSymbolicLearning