Bayesian Inference in Recurrent Explicit Duration Switching Linear Dynamical Systems

TL;DR

A novel model called REDSLDS is proposed that incorporates recurrent explicit duration variables into the rSLDS model and an inference and learning scheme that involves the use of Polya-gamma augmentation is proposed.

Abstract

In this paper, we propose a novel model called Recurrent Explicit Duration Switching Linear Dynamical Systems (REDSLDS) that incorporates recurrent explicit duration variables into the rSLDS model. We also propose an inference and learning scheme that involves the use of Pólya-gamma augmentation. We demonstrate the improved segmentation capabilities of our model on three benchmark datasets, including two quantitative datasets and one qualitative dataset.

Figures (9)

• Figure 1: The graphical model of REDSLDS
• Figure 2: The visualization of stick-breaking.
• Figure 3: NASCAR$^{\circledR}$ segmentations obtained for the $S \in \{5, 10, 15, 20\}$. We can observe that REDSLDS tends to have smoother and more coherent segmentations. However, it tends to ignore the second "straight-on" state.
• Figure 4: Trajectory of NASCAR$^{\circledR}$ used to test the models.
• Figure 5: Sample segmentations of the dance bee data set. As we can see, rSLDS tends to over-segment the bee movement. It especially struggles with "waggle" movement.