One interesting thing about the ARC competition is that it serves to highlight how people who use deep learning often have little idea of what deep learning actually does, and when they should be using it or not
DL is applicable when you're doing *pattern recognition*: when you have data that lies on a smooth manifold, along which samples can be interpolated. And you're going to need a dense sampling of your manifold as training data in order to fit a parametric approximation of it
Generalization in deep learning is interpolation along a latent manifold (or rather a learned approximation of it). It has little to do with your model itself and everything to do with the natural organization of your data
Differentiability & minibatch SGD are the strengths of DL: besides making the learning practically tractable, the smoothness & continuity of the function & the incrementality of its fitting work great to learn to approximate latent manifold. But its strengths are also its limits
The whole setup breaks down when you are no longer doing pattern recognition -- when you no longer have a latent manifold (any kind of discrete problem) or no longer have a dense sampling of it. Or when your manifold changes over time.
This isn't complicated
