Scaling Vision with Sparse Mixture of Experts

TL;DR

This paper introduces a sparse Vision MoE model that scales efficiently, achieves competitive accuracy with less compute, and offers adaptive inference, demonstrating significant potential for large-scale vision tasks.

Contribution

The paper presents a novel sparse Vision Transformer (V-MoE) that matches dense models' performance with reduced compute and introduces an adaptive routing algorithm for flexible inference.

Findings

V-MoE achieves state-of-the-art accuracy on ImageNet with less compute.

The adaptive routing allows smooth trade-offs between performance and compute.

A 15B parameter V-MoE attains 90.35% accuracy on ImageNet.

Abstract

Sparsely-gated Mixture of Experts networks (MoEs) have demonstrated excellent scalability in Natural Language Processing. In Computer Vision, however, almost all performant networks are "dense", that is, every input is processed by every parameter. We present a Vision MoE (V-MoE), a sparse version of the Vision Transformer, that is scalable and competitive with the largest dense networks. When applied to image recognition, V-MoE matches the performance of state-of-the-art networks, while requiring as little as half of the compute at inference time. Further, we propose an extension to the routing algorithm that can prioritize subsets of each input across the entire batch, leading to adaptive per-image compute. This allows V-MoE to trade-off performance and compute smoothly at test-time. Finally, we demonstrate the potential of V-MoE to scale vision models, and train a 15B parameter model that attains 90.35% on ImageNet.