Exploring the potential of closed-form continuous-time neural networks for physical process modelling

Closed-form continuous time neural networks

In order to model physical dynamical processes, differential equations can be used. These equations may be solved numerically but the computational cost increases as processes become more complex. In a novel approach, dynamical process are modelled using closed-form continuous depth artificial neural networks. On a variety of sequence modelling tasks, including the recognition of human actions and steering in autonomous vehicles, improved efficiency is demonstrated in both training and inference.