On Stacking Probabilistic Temporal Models with Bidirectional Information Flow

We discuss hierarchical combinations of probabilistic models where the upper layer is crafted for predicting time-series data. The combination of models makes the naı̈ve Bayes assumption, stating that the latent variables of the models are independent given the time-indexed label variables. In this setting an additional independence assumption between time steps and mildly inconsistent results are often accepted to make inference computationally feasible. We discuss how the application of approximate inference to the practically intractable joint model instead, shifts the need for these simplifications from model design time to inference time, and the application of loopy belief propagation to the joint model realizes bidirectional communication between models during inference. A first empirical evaluation of the proposed architecture on an activity recognition task demonstrates the benefits of the layered architecture and examines the effects of bidirectional information flow.