Learned Factor Graphs for Inference From Stationary Time Sequences

The design of methods for inference from time sequences has traditionally relied on statistical models that describe the relation between a latent desired sequence and observed one. A broad family model-based algorithms have been derived to carry out at controllable complexity using recursive computations over factor graph representing underlying distribution. An alternative model-agnostic approach utilizes machine learning (ML) methods. Here we propose framework combines data-driven ML tools stationary sequences. In proposed approach, neural networks are developed separately learn specific components describing distribution sequence, rather than complete task. By exploiting properties this distribution, resulting can be applied varying temporal duration. Learned graphs realized compact trainable small training sets, or alternatively, used improve upon existing deep systems. We present an algorithm based learned graphs, which learns implement sum-product scheme labeled data, different lengths. Our experimental results demonstrate ability sets accurate sleep stage detection Sleep-EDF dataset, as well symbol in digital communications with unknown channels.