Neural CDEs as Correctors for Learned Time Series Models
Researchers have developed a new Predictor-Corrector framework to enhance the accuracy of learned time series models. This framework utilizes a neural controlled differential equation as a Corrector to mitigate error accumulation in multi-step forecasts generated by a Predictor model. The Corrector is designed to handle irregularly sampled data and is compatible with various Predictor types, including neural ODEs and ContiFormer. Experiments across synthetic, physics-based, and real-world datasets demonstrate consistent improvements in forecasting performance, highlighting the framework's predictor-agnostic utility. AI
IMPACT Introduces a novel framework to improve the accuracy of AI-driven time series forecasting, potentially benefiting applications in scientific modeling and prediction.