Paying for your Lunch

There are a bunch of No Free Lunch proofs, but the canonical one is No free lunch theorems for search by Wolpert and Macready
The short version is that no machine learning algorithm is better on all problems.
One good thing about deep nets is that they support you paying for your lunch.
It should also be noted that the support vector machine framework also supports you paying for you lunch.
They both give you a set of tools to use to understand your data and to build systems that learn about that data, and give useful answers.
So with deep nets, if you have a vision problem, and you want to take advantage of things like receptive fields, you can build your initial layers so that they are connected in receptive fields.
If you want to build a Go game, you can just try to build the evaluation function in the deep net, and use standard game playing algorithms like mini-max from that.
You could also use an adversarial technique, so that the net plays a version of itself, so it doesn't need a real player to compete against.
You can use critics, based on the net, to present boundary conditions so that the system can focus on learning the important aspects of the problem.