Scalable spatiotemporal prediction with Bayesian neural fields.

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Spatiotemporal datasets are crucial across various fields like environmental monitoring and public health.
• Increasing data scale necessitates advanced statistical methods for complex dynamics and scalability.
• Existing methods often struggle with the high dimensionality and intricate patterns in modern spatiotemporal data.

Purpose of the Study:

• To introduce the Bayesian Neural Field (BAYESNF), a versatile statistical model for spatiotemporal data analysis.
• To enable accurate forecasting, interpolation, and variography on large-scale datasets.
• To provide robust uncertainty quantification for predictive modeling.

Main Methods:

• BAYESNF integrates deep neural networks for function estimation with hierarchical Bayesian inference.
• The model infers rich spatiotemporal probability distributions.
• Utilizes JAX for efficient computation on GPU and TPU accelerators.

Main Results:

• BAYESNF demonstrates improved prediction performance on climate and public health datasets.
• The model effectively handles datasets with tens to hundreds of thousands of measurements.
• Outperforms prominent baseline methods in prediction tasks.

Conclusions:

• BAYESNF offers a scalable and flexible solution for analyzing complex spatiotemporal data.
• The model provides reliable uncertainty quantification, crucial for decision-making.
• An open-source software package is available, facilitating broader adoption and research.