Re: Can 'backpropagation of error' work in humans?

Backpropagation is not generally considered neurally plausible for two
reasons:

1) As you mention, it would require that the "correct answer" be provided. 
Thus, when learning to crawl, there would have to be some teacher telling
the brain, "When you see and feel this, you should send these signals to
your limbs."  Clearly, there is no such teacher for pre-linguistic
processes.  In higher-level processes, such as language learning, there may
literally be a teacher: another person who tells you what the right answer
is.

2) Backpropagation requires information to travel "upstream", backward
along axons.  While this can be done in a petri dish, it does not happen in
the brain -- there is no mechanism to generate an action potential at the
synapse end of the axon.

There are less powerful, more realistic machine learning methods.  One
category of methods, reinforcement learning, does not give the learner the
correct answer, but does provide feedback on how good the learner's answer
was.  This is plausible as pleasure and pain reinforcement.  If an infant
felt more pleasure when she moved faster, her brain might adjust weights
(synapse strengths) to strengthen behavior patterns that cause such
movement, e.g., crawling.