[raindeer3]

To me this is a strangely written paper although the results might be some what interesting. They just optimize over a different set of functions than you normally do in NN training and find solutions to various problems. That they claim that the weights "are the solution" in normal NN training is wrong. Of course the weights AND the architecture are both important since they define the set of functions you search over. That you can find solutions by only searching over architecture is not surprising since the architecture is the combination of different operations.