The LICORS Cabinet: Nonparametric Algorithms for Spatio-temporal Prediction

For the task of unsupervised spatio-temporal forecasting (e.g., learning to predict video data without labels), we propose two new nonparametric predictive state algorithms, Moonshine and One Hundred Proof. The algorithms are conceptually simple and make few assumptions on the underlying spatio-temporal process yet have strong predictive performance and provide predictive distributions over spatio-temporal data. The latter property allows for likelihood estimation under the models, for classification and other probabilistic inference.