I'd be reluctant to trust anything serious, like a control system that
mattered, to an algorithm whose corner cases are undefined and
probably not testable.
We humans, _are_ neural nets.
Indeed we screw up plenty of times, but you sttill trust the train driver,
or bus driver, although you very well know accidents happen because
of human errors.
The neural nets (humans) also prevent accidents by not 'logically'
(as in and or xor) behaving sometimes.
Is not that contest with self-steering cars for a million $ prize
written out by DOD also using neural nets? Not sure.
Lemme google
http://www.neurosolutions.com/resources/darpa.html