Mode-Marginals: Expressing Uncertainty via Diverse M-Best Solutions

Representing uncertainty in predictions made by complex probabilistic models is often crucial but computationally challenging. There are a number of useful probabilistic models where computing the most probably assignment (or MAP) is easy but finding the marginal probability of variables is hard. In this paper, we present a novel representation of uncertainty in discrete probabilistic models, that we call Mode-Marginals. Mode-Marginals contain both max-marginals and marginals as special cases and express the entire spectrum in between the two. Computing mode-marginals involves performing MAP computation on deterministic perturbations to the unnormalized probability function, as introduced by Batra et al. [1]. We evaluate our method on a challenging computer vision application – human pose estimation, and demonstrate that mode-marginals provide a richer representation of uncertainty than max-marginals, leading to improved performance.