Neural Differential Equations for Time Series

Overview

Time series are often sampled at non-uniform intervals, and observations can contain gaps or noise. Neural Differential Equations (NDEs) model such data in continuous time, which lets them handle uneven sampling and missing values. Their continuous latent dynamics complement discrete-step sequence models, and domain knowledge can be built into the structure of the differential equation.

The main NDE families are Neural Ordinary Differential Equations (NODEs), Neural Controlled Differential Equations (NCDEs), and Neural Stochastic Differential Equations (NSDEs). Code examples use PyTorch, torchcde, and torchsde for tasks such as interpolation, representation learning, and classification.