Learning Latent Dynamics via Invariant Decomposition and (Spatio-)Temporal Transformers

We propose a method for learning dynamical systems from high-dimensional empirical data that combines variational autoencoders and (spatio-)temporal attention within a framework designed to enforce certain scientifically-motivated invariances. We focus on the setting in which data are available from multiple different instances of a system whose underlying dynamical model is entirely unknown at the outset. The approach rests on a separation into an instance-specific encoding (capturing initial conditions, constants etc.) and a latent dynamics model that is itself universal across all instances/realizations of the system. The separation is achieved in an automated, data-driven manner and only empirical data are required as inputs to the model. The approach allows effective inference of system behaviour at any continuous time but does not require an explicit neural ODE formulation, which makes it efficient and highly scalable. We study behaviour through simple theoretical analyses and extensive experiments on synthetic and real-world datasets. The latter investigate learning the dynamics of complex systems based on finite data and show that the proposed approach can outperform state-of-the-art neural-dynamical models. We study also more general inductive bias in the context of transfer to data obtained under entirely novel system interventions. Overall, our results provide a promising new framework for efficiently learning dynamical models from heterogeneous data with potential applications in a wide range of fields including physics, medicine, biology and engineering.