Spatial Mixture-of-Experts
This addresses the need for better spatial modeling in tasks like weather prediction, offering a novel method for capturing fine-grained structure, though it is incremental in building on mixture-of-experts concepts.
The paper tackles the problem of neural networks not leveraging spatial dependencies in data like weather or images, which violates assumptions like translation equivariance, by introducing the Spatial Mixture-of-Experts (SMoE) layer that learns spatial structure and routes experts finely; it shows strong results, setting new state-of-the-art for medium-range weather prediction and post-processing ensemble weather forecasts.
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.