Keywords: sum-product networks, probabilistic models, clustering, machine learning
Abstract: Recent research highlights the significance of incorporating density modeling into clustering procedures. While the Sum-Product Networks' ability to compactly represent mixture models has been long noticed, their potential for modal clustering remains largely unexplored. This paper explores the use of Gaussian Sum-Product Networks for semi-parametric density-based clustering via mode association. To associate points to modes, we make use of a recently developed efficient EM-style algorithm. We perform image segmentation experiments to evaluate the (dis)advantages of modal clustering using such models.
Submission Number: 4
Loading