Improving SLDS Performance Using Explicit Duration Variables with Infinite Support

Switching Linear Dynamical Systems (SLDS) are probabilistic graphical models used both for self-supervised segmentation and dimensionality reduction. Despite their modeling capabilities, SLDS are particularly hard to train. They oftentimes over-segment the timeseries or completely ignore some of the states, reducing the usefulness of the acquired segmentation. To improve the segmentation in Switching Linear Dynamical Systems, we introduce explicit-duration variables with infinite support. We extend the Beam Sampling algorithm to perform the efficient inference allowing for a duration distribution with infinite support. We conduct experiments on three benchmarks (two already prevalent in the state-space model literature and one demonstrating behavior in a sparse setting) that test the correctness and efficiency of our solution.