Spatial Mixture-of-Experts

TL;DR

The paper introduces the Spatial Mixture-of-Experts layer, a novel neural network component that captures spatial dependencies and fine-grained structure, improving performance on weather prediction and related tasks.

Contribution

It proposes the SMoE layer with new training techniques to effectively learn spatial structures, addressing limitations of existing models in capturing locality.

Findings

Achieves state-of-the-art results in medium-range weather prediction.

Demonstrates strong performance on ensemble weather forecast post-processing.

Introduces effective training methods for spatial mixture-of-experts layers.

Abstract

Many data have an underlying dependence on spatial location; it may be weather on the Earth, a simulation on a mesh, or a registered image. Yet this feature is rarely taken advantage of, and violates common assumptions made by many neural network layers, such as translation equivariance. Further, many works that do incorporate locality fail to capture fine-grained structure. To address this, we introduce the Spatial Mixture-of-Experts (SMoE) layer, a sparsely-gated layer that learns spatial structure in the input domain and routes experts at a fine-grained level to utilize it. We also develop new techniques to train SMoEs, including a self-supervised routing loss and damping expert errors. Finally, we show strong results for SMoEs on numerous tasks, and set new state-of-the-art results for medium-range weather prediction and post-processing ensemble weather forecasts.