>Can you honestly say with a straight face that there is no theory behind why convnets are the right choice?
There is theory of convnets, but it's more of a post-factum analysis, really, and I'm not sure it's behind much of anything at this point. Researchers use convnets because they work in ways that work for them.
Theory: 109 citations.
https://www.semanticscholar.org/paper/A-Mathematical-Theory-...
Practice: 42,551 citations.
https://www.semanticscholar.org/paper/ImageNet-Classificatio...