Abstract

Graphical techniques for modeling the of random variables have been explored in a variety of different areas including statistics, statistical physics, artificial intelligence (AI), speech recognition, processing, genetics. Formalisms for manipulating these models have been developed relatively independently in these research communities. In this paper we e hidden Markov models and related graphical structures within the general framework of directed acyclic graphs In particular we show that there exists a general methodology for deriving an exact inference method for any that this algorithm produces the well-known forward-backward and Viterbi algorithms as special cases, and that the complexity of inference can be characterized for general families of DAGs including HMMs which are augmented with non-local dependencies. Applications to problems such as speech recognition and biological sequence modeling . are discussed. 1 Introduction A directed acyclic graph is a graphical specificat...