Variational Non-Parametric Modelling of Spatiotemporal Signals

In this study, we explore Gaussian Process mixing models that describe complex observed spatiotemporal signals as transformations of simpler unobserved ones. Our main contribution is the development of several new generative models. Building upon current methods in the literature, we construct a variational inference scheme. We then use it to devise a series of variational mixing models that incorporate different structural assumptions about the generative process, including spatial sparsity, lowrank spatiotemporal covariance and some others.