Our Bayesian Deep Learning (BDL) climate surrogate model outperforms previous spatiotemporal models and estimates prediction uncertainty [1] (Figure 1).
Model explanation [2] (Figure 2) indicates the right features are captured by the model.
Significance and Impact
Advancing the state-of-the-art AI methods of DOE’s climate and earth system modeling research. Model and data uncertainty quantification helps decision making to mitigate climate-induced hazards to people, property, and infrastructure. Explanation and visualization validate model’s success is due to learning the correct signals from data.
With the Bayesian Deep Learning and Explainable Artificial Intelligence
Figure 1. Prediction performance. BDL: Bayesian Deep Learning
A Bayesian Deep Learning Approach to Near-Term Climate Prediction. Journal of Advances in Modeling Earth Systems. Accepted
Feature Importance in a Deep Learning Climate Emulator. ICLR 2021 Workshop on Modeling Oceans and Climate Change.
Figure 2. Model Explanation from pixel-wise contribution to the prediction
Technical Approach
Post-hoc local explanation methods, i.e., feature importance methods for understanding a deep learning emulator of climate data.
Bayesian deep learning models for emulating sea surface temperatures
Stein variational gradient descent method for efficient posterior analysis.