Probabilistic Graphical Models for Crowdsourcing and Turbulence

Abstract

Graphical models provide a useful framework and formalism from which to modeland solve problems involving random processes. We demonstrate the versatility and usefulness of graphical models on two problems, one involving crowdsourcing and one involving turbulence. In crowdsourcing, we consider the problem of inferring true labels from a set of crowdsourced annotations. We design generative models for the crowdsourced annotations involving as latent variables the worker reliability, the structure of the labels, and the ground truth labels. Furthermore, we design an effective inference algorithm to infer the latent variables. In turbulence, we consider the problem of modeling the mixing distribution of homogeneous isotropic passive scalar turbulence. We consider models specifying the conditional distribution of a coarse grained node given its adjacent coarse grained nodes. In particular, we demonstrate the effectiveness of a higher order moments based extension of the Gaussian distribution.