Jeff A. Bilmes

Integrative annotation of chromatin elements from ENCODE data

The ENCODE Project has generated a wealth of experimental information mapping diverse chromatin properties in several human cell lines. Although each such data track is independently informative toward the annotation of regulatory elements, their interrelations contain much richer information for the systematic annotation of regulatory elements. To uncover these interrelations and to generate an interpretable summary of the massive datasets of the ENCODE Project, we apply unsupervised learning methodologies, converting dozens of chromatin datasets into discrete annotation maps of regulatory regions and other chromatin elements across the human genome. These methods rediscover and summarize diverse aspects of chromatin architecture, elucidate the interplay between chromatin activity and RNA transcription, and reveal that a large proportion of the genome lies in a quiescent state, even across multiple cell types. The resulting annotation of non-coding regulatory elements correlate strongly with mammalian evolutionary constraint, and provide an unbiased approach for evaluating metrics of evolutionary constraint in human. Lastly, we use the regulatory annotations to revisit previously uncharacterized disease-associated loci, resulting in focused, testable hypotheses through the lens of the chromatin landscape.

Factored language models and generalized parallel backoff

We introduce factored language models (FLMs) and generalized parallel backoff (GPB). An FLM represents words as bundles of features (e.g., morphological classes, stems, data-driven clusters, etc.), and induces a probability model covering sequences of bundles rather than just words. GPB extends standard backoff to general conditional probability tables where variables might be heterogeneous types, where no obvious natural (temporal) backoff order exists, and where multiple dynamic backoff strategies are allowed. These methodologies were implemented during the JHU 2002 workshop as extensions to the SRI language modeling toolkit. This paper provides initial perplexity results on both CallHome Arabic and on Penn Treebank Wall Street Journal articles. Significantly, FLMs with GPB can produce bigrams with significantly lower perplexity, sometimes lower than highly-optimized baseline trigrams. In a multi-pass speech recognition context, where bigrams are used to create first-pass bigram lattices or N-best lists, these results are highly relevant.

The graphical models toolkit: An open source software system for speech and time-series processing

This paper describes the Graphical Models Toolkit (GMTK), an open source, publically available toolkit for developing graphical-model based speech recognition and general time series systems. Graphical models are a flexible, concise, and expressive probabilistic modeling framework with which one may rapidly specify a vast collection of statistical models. This paper begins with a brief description of the representational and computational aspects of the framework. Following that is a detailed description of GMTKs features, including a language for specifying structures and probability distributions, logarithmic space exact training and decoding procedures, the concept of switching parents, and a generalized EM training method which allows arbitrary sub-Gaussian parameter tying. Taken together, these features endow GMTK with a degree of expressiveness and functionality that significantly complements other publically available packages. GMTK was recently used in the 2001 Johns Hopkins Summer Workshop, and experimental results are described in detail both herein and in a companion paper.

What HMMs Can Do

Since their inception almost fifty years ago, hidden Markov models (HMMs) have have become the predominant methodology for automatic speech recognition (ASR) systems---today, most state-of-the-art speech systems are HMM-based. There have been a number of ways to explain HMMs and to list their capabilities, each of these ways having both advantages and disadvantages. In an effort to better understand what HMMs can do, this tutorial article analyzes HMMs by exploring a definition of HMMs in terms of random variables and conditional independence assumptions. We prefer this definition as it allows us to reason more throughly about the capabilities of HMMs. In particular, it is possible to deduce that there are, in theory at least, no limitations to the class of probability distributions representable by HMMs. This paper concludes that, in search of a model to supersede the HMM (say for ASR), rather than trying to correct for HMM limitations in the general case, new models should be found based on their potential for better parsimony, computational requirements, and noise insensitivity.

Transmembrane Topology and Signal Peptide Prediction Using Dynamic Bayesian Networks

Hidden Markov models (HMMs) have been successfully applied to the tasks of transmembrane protein topology prediction and signal peptide prediction. In this paper we expand upon this work by making use of the more powerful class of dynamic Bayesian networks (DBNs). Our model, Philius, is inspired by a previously published HMM, Phobius, and combines a signal peptide submodel with a transmembrane submodel. We introduce a two-stage DBN decoder that combines the power of posterior decoding with the grammar constraints of Viterbi-style decoding. Philius also provides protein type, segment, and topology confidence metrics to aid in the interpretation of the predictions. We report a relative improvement of 13% over Phobius in full-topology prediction accuracy on transmembrane proteins, and a sensitivity and specificity of 0.96 in detecting signal peptides. We also show that our confidence metrics correlate well with the observed precision. In addition, we have made predictions on all 6.3 million proteins in the Yeast Resource Center (YRC) database. This large-scale study provides an overall picture of the relative numbers of proteins that include a signal-peptide and/or one or more transmembrane segments as well as a valuable resource for the scientific community. All DBNs are implemented using the Graphical Models Toolkit. Source code for the models described here is available at http://noble.gs.washington.edu/proj/philius. A Philius Web server is available at http://www.yeastrc.org/philius, and the predictions on the YRC database are available at http://www.yeastrc.org/pdr.

MVA Processing of Speech Features

In this paper, we investigate a technique consisting of mean subtraction, variance normalization and time sequence filtering. Unlike other techniques, it applies auto-regression moving-average (ARMA) filtering directly in the cepstral domain. We call this technique mean subtraction, variance normalization, and ARMA filtering (MVA) post-processing, and speech features with MVA post-processing are called MVA features. Overall, compared to raw features without post-processing, MVA features achieve an error rate reduction of 45% on matched tasks and 65% on mismatched tasks on the Aurora 2.0 noisy speech database, and an average 57% error reduction on the Aurora 3.0 database. These improvements are comparable to the results of much more complicated techniques even though MVA is relatively simple and requires practically no additional computational cost. In this paper, in addition to describing MVA processing, we also present a novel analysis of the distortion of mel-frequency cepstral coefficients and the log energy in the presence of different types of noise. The effectiveness of MVA is extensively investigated with respect to several variations: the configurations used to extract and the type of raw features, the domains where MVA is applied, the filters that are used, the ARMA filter orders, and the causality of the normalization process. Specifically, it is argued and demonstrated that MVA works better when applied to the zeroth-order cepstral coefficient than to log energy, that MVA works better in the cepstral domain, that an ARMA filter is better than either a designed finite impulse response filter or a data-driven filter, and that a five-tap ARMA filter is sufficient to achieve good performance in a variety of settings. We also investigate and evaluate a multi-domain MVA generalization

Graphical model architectures for speech recognition

This article discusses the foundations of the use of graphical models for speech recognition as presented in J. R. Deller et al. (1993), X. D. Huang et al. (2001), F. Jelinek (19970, L. R. Rabiner and B. -H. Juang (1993) and S. Young et al. (1990) giving detailed accounts of some of the more successful cases. Our discussion employs dynamic Bayesian networks (DBNs) and a DBN extension using the Graphical Model Toolkits (GMTKs) basic template, a dynamic graphical model representation that is more suitable for speech and language systems. While this article concentrates on speech recognition, it should be noted that many of the ideas presented here are also applicable to natural language processing and general time-series analysis.

Graphical Models and Automatic Speech Recognition

Graphical models provide a promising paradigm to study both existing and novel techniques for automatic speech recognition. This paper first provides a brief overview of graphical models and their uses as statistical models. It is then shown that the statistical assumptions behind many pattern recognition techniques commonly used as part of a speech recognition system can be described by a graph – this includes Gaussian distributions, mixture models, decision trees, factor analysis, principle component analysis, linear discriminant analysis, and hidden Markov models. Moreover, this paper shows that many advanced models for speech recognition and language processing can also be simply described by a graph, including many at the acoustic-, pronunciation-, and language-modeling levels. A number of speech recognition techniques born directly out of the graphical-models paradigm are also surveyed. Additionally, this paper includes a novel graphical analysis regarding why derivative (or delta) features improve hidden Markov model-based speech recognition by improving structural discriminability. It also includes an example where a graph can be used to represent language model smoothing constraints. As will be seen, the space of models describable by a graph is quite large. A thorough exploration of this space should yield techniques that ultimately will supersede the hidden Markov model.

Submodularity beyond submodular energies: Coupling edges in graph cuts

We propose a new family of non-submodular global energy functions that still use submodularity internally to couple edges in a graph cut. We show it is possible to develop an efficient approximation algorithm that, thanks to the internal submodularity, can use standard graph cuts as a subroutine. We demonstrate the advantages of edge coupling in a natural setting, namely image segmentation. In particular, for fine-structured objects and objects with shading variation, our structured edge coupling leads to significant improvements over standard approaches.

Statistical Models for Empirical Search-Based Performance Tuning

Achieving peak performance from the computational kernels that dominate application performance often requires extensive machine-dependent tuning by hand. Automatic tuning systems have emerged in response, and they typically operate by (1) generating a large number of possible, reasonable implementations of a kernel, and (2) selecting the fastest implementation by a combination of heuristic modeling, heuristic pruning, and empirical search (i.e. actually running the code). This paper presents quantitative data that motivate the development of such a search-based system, using dense matrix multiply as a case study. The statistical distributions of performance within spaces of reasonable implementations, when observed on a variety of hardware platforms, lead us to pose and address two general problems which arise during the search process. First, we develop a heuristic for stopping an exhaustive compiletime search early if a near-optimal implementation is found. Secondly, we show how to construct run-time decision rules, based on run-time inputs, for selecting from among a subset of the best implementations when the space of inputs can be described by continuously varying features. We address both problems by using statistical modeling techniques that exploit the large amount of performance data collected during the search. We demonstrate these methods on actual performance data collected by the PHiPAC tuning system for dense matrix multiply. We close with a survey of recent projects that use or otherwise advocate an empirical search-based approach to code generation and algorithm selection, whether at the level of computational kernels, compiler and run-time systems, or problem-solving environments. Collectively, these efforts suggest a number of possible software architectures for constructing platform-adapted libraries and applications.