This project is significantly enhancing available software infrastructure, the graphical models toolkit (GMTK), for graphical-model based time series modeling. While GMTK is currently widely used for speech recognition, the extensions being performed, however, are optimized not just for speech recognition but for all time-series applications. New infrastructure is also being developed that significantly enhances GMTK?s abilities, speed, documentation, source code availability, and pedagogical structure. This work will give to both the student and the researcher an enormous number of dynamic graphical model facilities. With its new features, GMTK will be able to perform computationally difficult time-series processing on extremely large and diverse data sets.
The main goal in this project is to significantly improve and then make publicly available for free software infrastructure for statistical modeling, specifically, GMTK (the Graphical Models ToolKit). The Graphical Models Toolkit (GMTK) is an open source, publicly available toolkit that allows a researcher or scientist to rapidly prototype, test, learn, and then use statistical models under the framework of dynamic graphical models (DGMs) and dynamic Bayesian networks (DBNs). GMTK can be used for applications and research in speech and language processing, bioinformatics, activity recognition, econometrics, or any type of time series or sequential data application. GMTK has many features, including exact and approximate inference; a large variety of built-in factors including dense, sparse, and deterministic conditional probability tables, native support for ARPA backoff-based factors and factored language models (FLMs), parameter sharing, gamma and beta distributions, dense and sparse Gaussian factors, heterogeneous mixtures, deep neural network factors and CPTs, and time-inhomogeneous trellis factors; arbitrary order embedded Markov chains; a GUI-based graph viewer; flexible feature-file support and processing tools (supporting pfiles, HTK files, ASCII/binary, and HDF5 files); parameter learning algorithms; rich graph triangulation engine and exact and approximate dynamic inference procedures; and both offline and streaming online inference methods that can be used for both parameter learning and prediction. More information is available in the documentation. All in all, GMTK offers a flexible, concise, and expressive probabilistic modeling framework with which one may rapidly specify a vast collection of temporal statistical models. GMTK is meant as a toolkit (meaning there are a set of top-level programs that you call from the command line and which are often called from scripts or other programs), rather than a library (where you have an API which you can call from your own front-end C++ code). GMTK it has its own graph specification language and tokenizer/parser, many commands and command-line options, and the idea is that the user does most everything via that mechanism. Of course, given the GMTK source code (which is available online), the core GMTK routines for inference can easily be linked to and it is possible to make ones own front-end program. GMTK currently is available to download for free from the web site http://melodi.ee.washington.edu/gmtk/ and there is also a 650 page documentation that describes details of the toolkit as well. The tutorials includes over 100 example working models, each with its own brief documentation, and also includes a simple but complete step-by-step tutorial on how to develop, test, and then use a model.
