You're absolutely correct - the intention of this certainly isn't to somehow simplify the process of building neural nets, which, if you're just running things locally at least, is fairly straightforward. The intention of the drag and drop interface is to give a clear and consistent way to visualize your neural net architecture. It also lends itself to a nice way to display per layer outputs.
i say this because most the people i know that are doing graduate studies focused on general relativity are in theoretical physics. we're not at the point, technologically, where experimental physics can cover the full range of what general relativity describes. it would be nice to have a black hole in a lab that is not an analogue, though. and one day, we probably will. so, i agree that eventually it will be on par with electromagnetism, but for now it describes a lot that is still very theoretical.
But EM as a subject, as its usually taught, is still theoretical. Experimental EM would be better known as either electronics or statistics. You make it sound like theoretical means lacking experimental support, perhaps that is what the word means in, say, politics, but certainly not in physics.
OR you can simply project the fourth dimension into 2D, much like how we project 3D pictures onto 2D screens. Here's an example http://jakebian.github.io/hypercanvas.js/
What would your response be to a child who said something similar when you drew a cube?
It does take practise to be able to visualise 4d objects, and I agree that 2d projection example isn't helpful. I find it interesting to think of the 3d -> 2d problem that some people have, and extend it to 4d.
I mean that there is no description on how these projections are even created.
A child would recognize a cube since we naturally see 2d projections of 3d objects as it is. Likewise a 2 dimensional being could recognize a 1d projection of a 2d object (looking at the 2d object on it's side.) But would a 3d object projected to 1 dimension make any sense?
