Learning nonparametric ordinary differential equations from noisy data.

Area of Science:

• Machine Learning
• Dynamical Systems
• Applied Mathematics

Background:

• Learning nonparametric systems of Ordinary Differential Equations (ODEs) from noisy data is a challenging emerging topic in machine learning.
• Reproducing Kernel Hilbert Spaces (RKHS) provide a robust theoretical framework for defining ODE candidates with guaranteed existence and uniqueness of solutions.

Purpose of the Study:

• To develop a novel method for learning nonparametric ODEs from noisy data.
• To leverage RKHS theory for defining and learning ODE systems.
• To demonstrate the method's effectiveness on benchmark systems and a biological prediction task.

Main Methods:

• Utilizing RKHS to define candidate ODEs ensuring unique solutions.
• Formulating the learning problem as a constrained optimization within an RKHS.
• Proposing an iterative penalty method employing the Representer theorem and Euler approximations for numerical solutions.

Main Results:

• Proving a generalization bound for the distance between the true ODE and its learned estimator.
• Achieving competitive performance on the FitzHugh-Nagumo oscillator and Lorenz system.
• Demonstrating successful prediction of Amyloid levels in aging subjects' cortex.

Conclusions:

• The proposed RKHS-based penalty method offers an effective approach for learning nonparametric ODEs from noisy data.
• The method shows strong performance across diverse applications, including complex dynamical systems and biomedical predictions.
• This work contributes a valuable tool for analyzing and predicting dynamic processes in various scientific domains.