it would be helpful if we knew whether for almost all random functions from ℝⁿ→ℝ, changing any one element of the domain slightly changes the output a lot

then one might also prove (or disprove) the same thing for functions implementable by some classes neural networks