Unsupervised Learning is mainly good for Feature Detection

Unsupervised learning is useful for low-level wetware emulation ... feature detection, signal recognition, etc. 

However, most of what we call civilisation and social life, depend on higher-level constructs which are maintained by supervised learning in humans, so it is only reasonable that we should teach anthropomorphic machines using the same methods which we already apply to humans.

I can never understand people who believe it is more efficient to do otherwise - though I would be happy to consider proofs.

( Updated my vocabulary today, with 'feature' and 'blackboard' : for concepts which I am already quite familiar with. )