Scaling Vision with Sparse Mixture of Experts

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.

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Image Classification ImageNet V-MoE-H/14 (Last-5) Top 1 Accuracy 88.23% # 32
Number of params 2700M # 678
Image Classification ImageNet V-MoE-H/14 (Every-2) Top 1 Accuracy 88.36% # 30
Number of params 7200M # 680
Image Classification ImageNet VIT-H/14 Top 1 Accuracy 88.08% # 33
Number of params 656M # 661
Image Classification ImageNet V-MoE-L/16 (Every-2) Top 1 Accuracy 87.41% # 45
Number of params 3400M # 679
Few-Shot Image Classification ImageNet - 10-shot V-MoE-H/14 (Last-5) Top 1 Accuracy 80.1 # 3
Few-Shot Image Classification ImageNet - 10-shot VIT-H/14 Top 1 Accuracy 79.01 # 4
Few-Shot Image Classification ImageNet - 10-shot ViT-MoE-15B (Every-2) Top 1 Accuracy 84.29 # 1
Few-Shot Image Classification ImageNet - 10-shot V-MoE-H/14 (Every-2) Top 1 Accuracy 80.33 # 2
Few-Shot Image Classification ImageNet - 1-shot ViT-MoE-15B (Every-2) Top 1 Accuracy 68.66 # 1
Few-Shot Image Classification ImageNet - 1-shot V-MoE-L/16 (Every-2) Top 1 Accuracy 62.41 # 4
Few-Shot Image Classification ImageNet - 1-shot VIT-H/14 Top 1 Accuracy 62.34 # 5
Few-Shot Image Classification ImageNet - 1-shot V-MoE-H/14 (Last-5) Top 1 Accuracy 62.95 # 3
Few-Shot Image Classification ImageNet - 1-shot V-MoE-H/14 (Every-2) Top 1 Accuracy 63.38 # 2
Few-Shot Image Classification ImageNet - 5-shot ViT-MoE-15B (Every-2) Top 1 Accuracy 82.78 # 1
Few-Shot Image Classification ImageNet - 5-shot V-MoE-H/14 (Every-2) Top 1 Accuracy 78.21 # 2
Few-Shot Image Classification ImageNet - 5-shot V-MoE-H/14 (Last-5) Top 1 Accuracy 78.08 # 3
Few-Shot Image Classification ImageNet - 5-shot V-MoE-L/16 (Every-2) Top 1 Accuracy 77.1 # 4
Few-Shot Image Classification ImageNet - 5-shot VIT-H/14 Top 1 Accuracy 76.95 # 5
Image Classification JFT-300M V-MoE-H/14 (Last-5) prec@1 60.12 # 2
Image Classification JFT-300M VIT-H/14 prec@1 56.68 # 4
Image Classification JFT-300M V-MoE-H/14 (Every-2) prec@1 60.62 # 1
Image Classification JFT-300M V-MoE-L/16 (Every-2) prec@1 57.65 # 3

Methods