On: Why deep learning works even though it shouldn’t

[bananaface]

Doesn't the fact the network is discrete (floats have a maximum precision) mean this isn't actually the case? There's a finite number of states the net can be in, and one (or more) is best.