Comment by AstralStorm

Have you actually tried to use one of those networks on a decent real dataset? They're making common category errors. ImageNet roulette is an interesting fun example of what happens if you feed an overfitted network with real images. Such as misclassifying children. :)