Saw that deep-learning paper, thanks. As far as I can tell, they haven't been able to reverse engineer any of the connections in their model yet. So it may work but they don't understand why. If you have something I missed that would be interesting to hear about.
It points out how connections are developed. For example, in the finding I made in the previous comment
, the neural net would need to have picked out correlations in frequency space by swapping out various frequencies and doing comparisons. So, in this case, the net would need to have been designed to accommodate that.