15
1

Approximation Rates and VC-Dimension Bounds for (P)ReLU MLP Mixture of Experts

Abstract

Mixture-of-Experts (MoEs) can scale up beyond traditional deep learning models by employing a routing strategy in which each input is processed by a single "expert" deep learning model. This strategy allows us to scale up the number of parameters defining the MoE while maintaining sparse activation, i.e., MoEs only load a small number of their total parameters into GPU VRAM for the forward pass depending on the input. In this paper, we provide an approximation and learning-theoretic analysis of mixtures of expert MLPs with (P)ReLU activation functions. We first prove that for every error level ε>0\varepsilon>0 and every Lipschitz function f:[0,1]nRf:[0,1]^n\to \mathbb{R}, one can construct a MoMLP model (a Mixture-of-Experts comprising of (P)ReLU MLPs) which uniformly approximates ff to ε\varepsilon accuracy over [0,1]n[0,1]^n, while only requiring networks of O(ε1)\mathcal{O}(\varepsilon^{-1}) parameters to be loaded in memory. Additionally, we show that MoMLPs can generalize since the entire MoMLP model has a (finite) VC dimension of O~(Lmax{nL,JW})\tilde{O}(L\max\{nL,JW\}), if there are LL experts and each expert has a depth and width of JJ and WW, respectively.

View on arXiv
Comments on this paper