Solving Multiple Inference by Minimizing Expected Loss
Proceedings of the 10th International Conference on Probabilistic Graphical Models, PMLR 138:65-76, 2020.
Multiple Inference is the problem of finding multiple top solutions for an inference problem in a graphical model. It has been shown that it is beneficial for the top solutions to be diverse. However, existing methods, such as diverse M-Best and M-Modes, often rely on a hyper parameter in enforcing diversity. The optimal values of such parameters usually depend on the probability landscape of the graphical model and thus have to be tuned case by case via cross validation. This is not a desirable property. In this paper, we introduce a parameter-free method that directly minimizes the expected loss of each solution in finding multiple top solutions that have high oracle accuracy, and are automatically diverse. Empirical evaluations show that our method often have better performance than other competing methods.