Nick — Data Science freelancer in Nashville

Key problems in computer vision, algorithmic trading, cost optimization, social network analysis, and many other fields are represented with probabilistic graphical models (PGMs). Unfortunately, the key mathematical problem underlying PGM learning and prediction tasks, known as probabilistic inference, is computationally intractable. My PhD research and thesis advanced the state of the art in probabilistic inference algorithms for both classic and emerging application domains. My thesis developed principled methods for exploiting symmetry - meaning groups of indistinguishable objects - in graphical models. Two types of symmetry arise regularly in practice: first, a model with state space symmetry contains factors with groups of states that have identical values; second, a model with factor symmetry contains groups of factors that have identical factor tables. Although efficient inference algorithms for models with perfect symmetry have been well developed, most real problems do not contain perfect symmetry. Often, for example, a model with a symmetric substructure is perturbed by asymmetric evidence factors. Consequently, there is a great need for methods that accurately approximate the desired inference quantities using symmetric inference terms. While a few methods to address this problem previously existed, most are heuristic or addressed certain aspects of the problem.