One of thing I love telling the that always surprises people is that you can't build a deep learning model that can do math (at least using conventional layers).
I'm curious what approaches you're thinking about. When last looking into the matter I found some research in Neural Turing Machines, but they're so obscure I hadn't ever heard of them and assume they're not widely used.
While you could build a model to answer math questions for a set input space, these approaches break down once you expand beyond the input space.
Yeah, but since Neural networks are really function approximators, the farther you move away from the training input space, the higher the error rate will get. For multiplication it gets worse because layers are generally additive, so you'd need layers = largest input value to work.
Is that a thing? Looking it up I really only see a couple one off papers on mixing deep learning and finite state machines. Do you have examples or references to what you're talking about, or is it just a concept?