11
35

Depth Separations in Neural Networks: What is Actually Being Separated?

Abstract

Existing depth separation results for constant-depth networks essentially show that certain radial functions in Rd\mathbb{R}^d, which can be easily approximated with depth 33 networks, cannot be approximated by depth 22 networks, even up to constant accuracy, unless their size is exponential in dd. However, the functions used to demonstrate this are rapidly oscillating, with a Lipschitz parameter scaling polynomially with the dimension dd (or equivalently, by scaling the function, the hardness result applies to O(1)\mathcal{O}(1)-Lipschitz functions only when the target accuracy ϵ\epsilon is at most poly(1/d)\text{poly}(1/d)). In this paper, we study whether such depth separations might still hold in the natural setting of O(1)\mathcal{O}(1)-Lipschitz radial functions, when ϵ\epsilon does not scale with dd. Perhaps surprisingly, we show that the answer is negative: In contrast to the intuition suggested by previous work, it \emph{is} possible to approximate O(1)\mathcal{O}(1)-Lipschitz radial functions with depth 22, size poly(d)\text{poly}(d) networks, for every constant ϵ\epsilon. We complement it by showing that approximating such functions is also possible with depth 22, size poly(1/ϵ)\text{poly}(1/\epsilon) networks, for every constant dd. Finally, we show that it is not possible to have polynomial dependence in both d,1/ϵd,1/\epsilon simultaneously. Overall, our results indicate that in order to show depth separations for expressing O(1)\mathcal{O}(1)-Lipschitz functions with constant accuracy -- if at all possible -- one would need fundamentally different techniques than existing ones in the literature.

View on arXiv
Comments on this paper