Abstract
Interacting systems are prevalent in nature, from dynamical systems inphysics to complex societal dynamics. The interplay of components can give riseto complex behavior, which can often be explained using a simple model of thesystem's constituent parts. In this work, we introduce the neural relationalinference (NRI) model: an unsupervised model that learns to infer interactionswhile simultaneously learning the dynamics purely from observational data. Ourmodel takes the form of a variational auto-encoder, in which the latent coderepresents the underlying interaction graph and the reconstruction is based ongraph neural networks. In experiments on simulated physical systems, we showthat our NRI model can accurately recover ground-truth interactions in anunsupervised manner. We further demonstrate that we can find an interpretablestructure and predict complex dynamics in real motion capture and sportstracking data.