Conditional Flow Matching for Time Series Modelling

Learning dynamical systems from long trajectories is a challenging problem due to the complexity of the loss landscape. Inspired by conditional flow matching in generative modeling, we propose a new approach for training neural ODEs based on regressing vector fields of conditional probability paths defined per trajectory. Our Conditional Flow Matching for Time Series (CFM-TS) objective outperforms neural ODEs trained with the adjoint method on three simulated tasks, including a pendulum system where the neural ODE loss does not converge at all.