Continual Learning and Adaptation In Resource-Constrained Environments (CLAIRE)

Abstract

As climate-related extreme events continue to increase and impact the world, of particular importance are the threats posed to food, agriculture, and water (FAW) systems. Deep learning could benefit FAW systems for classification of threats and for forecasting potential future events given historical patterns. However, many FAW systems are faced with operational environments that are resource-constrained, which could present challenges in deploying deep learning models. Continual learning offers a way to overcome certain deployment challenges by enabling deep learning models that are more robust to data distribution changes, without the need for GPUs or off-line training. We describe a continual learning approach to forecasting extreme air quality events developed for the National Oceanic and Atmospheric Administration to provide operational air quality guidance to the Continental United States. We describe how this deep learning model is resilient to future data distribution changes by performing curriculum learning, and how it can be deployed as a continual learner, offering better predictive performance for resource-constrained environments.