Mixture-of-Experts Models in Vision: Routing, Optimization, and Generalization

TL;DR

This paper investigates the behavior of mixture-of-experts models in image classification, analyzing their performance, expert utilization, and generalization, and comparing them to dense models on CIFAR10.

Contribution

It provides a detailed analysis of MoE models in vision, including their generalization, curvature properties, and inference efficiency, which has been less explored compared to language models.

Findings

MoE models slightly outperform dense models in validation accuracy.

SoftMoE exhibits higher sharpness metrics than dense and SparseMoE.

Naive routing does not improve inference speed on modern hardware.

Abstract

Mixture-of-Experts (MoE) architectures enable conditional computation by routing inputs to multiple expert subnetworks and are often motivated as a mechanism for scaling large language models. In this project, we instead study MoE behavior in an image classification setting, focusing on predictive performance, expert utilization, and generalization. We compare dense, SoftMoE, and SparseMoE classifier heads on the CIFAR10 dataset under comparable model capacity. Both MoE variants achieve slightly higher validation accuracy than the dense baseline while maintaining balanced expert utilization through regularization, avoiding expert collapse. To analyze generalization, we compute Hessian-based sharpness metrics at convergence, including the largest eigenvalue and trace of the loss Hessian, evaluated on both training and test data. We find that SoftMoE exhibits higher sharpness by these metrics, while Dense and SparseMoE lie in a similar curvature regime, despite all models achieving comparable generalization performance. Complementary loss surface perturbation analyses reveal qualitative differences in non-local behavior under finite parameter perturbations between dense and MoE models, which help contextualize curvature-based measurements without directly explaining validation accuracy. We further evaluate empirical inference efficiency and show that naively implemented conditional routing does not yield inference speedups on modern hardware at this scale, highlighting the gap between theoretical and realized efficiency in sparse MoE models.