Latent State Inference in a Spatiotemporal Generative Model

Knowledge about the hidden factors that determine particular system dynamics is crucial for both explaining them and pursuing goal-directed interventions. Inferring these from time series data without supervision remains an open challenge. Here, we focus on spatiotemporal processes, including wave propagation weather dynamics, which assume universal causes (e.g. physics) apply throughout space time. A recently introduced DIstributed SpatioTemporal graph Artificial Neural network Architecture (DISTANA) used enhanced to learn such requiring fewer parameters achieving significantly more accurate predictions compared temporal convolutional neural networks other related approaches. We show DISTANA, when combined with a retrospective latent state inference principle called active tuning, can reliably derive location-respective causal factors. In current prediction benchmark, DISTANA infers our planet’s land-sea mask solely by observing temperature and, meanwhile, uses self inferred information improve its own future predictions.