Efficient Heuristic Search for M-Modes Inference
Proceedings of the 10th International Conference on Probabilistic Graphical Models, PMLR 138:77-88, 2020.
M-Modes is the problem of finding the top M locally optimal solutions of a graphical model, called modes. These modes provide geometric characterization of the energy landscape of a graphical model and lead to high quality solutions in structured prediction. It has been shown that any mode must be a local MAP within every subgraph of certain size. The state-of-the-art method is a search algorithm that explores subgraphs in a fixed ordering, uses each subgraph as a layer and searches for a consistent concatenation of local MAPs. We observe that for the M-Modes problem, different search orderings can lead to search spaces with dramatically different sizes, resulting in huge differences in performance. We formalize a metric measuring the quality of different orderings. We then formulate finding an optimized ordering as a shortest path problem, and introduce pruning criteria to speed up the search. Our empirical results show that using optimized orderings improves the efficiency of M-Modes search by up to orders of magnitude.