Performance Of Hidden Markov Models Research Articles

A survey of techniques for incremental learning of HMM parameters

The performance of Hidden Markov Models (HMMs) targeted for complex real-world applications are often degraded because they are designed a priori using limited training data and prior knowledge, and because the classification environment changes during operations. Incremental learning of new data sequences allows to adapt HMM parameters as new data becomes available, without having to retrain from the start on all accumulated training data. This paper presents a survey of techniques found in literature that are suitable for incremental learning of HMM parameters. These techniques are classified according to the objective function, optimization technique and target application, involving block-wise and symbol-wise learning of parameters. Convergence properties of these techniques are presented along with an analysis of time and memory complexity. In addition, the challenges faced when these techniques are applied to incremental learning is assessed for scenarios in which the new training data is limited and abundant. While the convergence rate and resource requirements are critical factors when incremental learning is performed through one pass over abundant stream of data, effective stopping criteria and management of validation sets are important when learning is performed through several iterations over limited data. In both cases managing the learning rate to integrate pre-existing knowledge and new data is crucial for maintaining a high level of performance. Finally, this paper underscores the need for empirical benchmarking studies among techniques presented in literature, and proposes several evaluation criteria based on non-parametric statistical testing to facilitate the selection of techniques given a particular application domain.

Performance of hidden Markov model and dynamic Bayesian network classifiers on handwritten Arabic word recognition

This paper presents a comparative study of two machine learning techniques for recognizing handwritten Arabic words, where hidden Markov models (HMMs) and dynamic Bayesian networks (DBNs) were evaluated. The work proposed is divided into three stages, namely preprocessing, feature extraction and classification. Preprocessing includes baseline estimation and normalization as well as segmentation. In the second stage, features are extracted from each of the normalized words, where a set of new features for handwritten Arabic words is proposed, based on a sliding window approach moving across the mirrored word image. The third stage is for classification and recognition, where machine learning is applied using HMMs and DBNs. In order to validate the techniques, extensive experiments were conducted using the IFN/ENIT database which contains 32,492 Arabic words. Experimental results and quantitative evaluations showed that HMM outperforms DBN in terms of higher recognition rate and lower complexity.

Modeling default risk via a hidden Markov model of multiple sequences

Default risk in commercial lending is one of the major concerns of the creditors. In this article, we introduce a new hidden Markov model (HMM) with multiple observable sequences (MHMM), assuming that all the observable sequences are driven by a common hidden sequence, and utilize it to analyze default data in a network of sectors. Efficient estimation method is then adopted to estimate the model parameters. To further illustrate the advantages of MHMM, we compare the hidden risk state process obtained by MHMM with that from the traditional HMMs using credit default data. We then consider two applications of our MHMM. The calculation of two important risk measures: Value-at-risk (VaR) and expected shortfall (ES) and the prediction of global risk state. We first compare the performance of MHMM and HMM in the calculation of VaR and ES in a portfolio of default-prone bonds. A logistic regression model is then considered for the prediction of global economic risk using our MHMM with default data. Numerical results indicate our model is effective for both applications.

Combined prediction of transmembrane topology and signal peptide of β-barrel proteins: Using a hidden Markov model and genetic algorithms

Background Hidden Markov models (HMMs) have been extensively used in computational molecular biology, for modelling protein and nucleic acid sequences. The design of the model architecture and the algorithms for parameter estimation and decoding are extremely important for improve the performance of HMM. In topology prediction of transmembrane β-barrels proteins (TMBs), the Baum–Welch algorithm is widely adapted for HMM training but usually leads to a sub-optimal model in practice. In addition, all the existing HMM-based predictors are only designed to model the transmembrane segment without a submodel to model the signal peptide (SP) for full-length sequences. It is not convenient for users to investigate the structures of full-length TMB sequences. Results We present here, an HMM that combine a transmembrane barrel submodel and an SP submodel for both topology and SP predictions. A new genetic algorithm (GA) is presented here to training the model, at the same time the Posterior–Viterbi algorithm is adopted for decoding. A dataset including 33 TMBs that is the most so far in literature are collected for model training and testing. Results of self-consistency and jackknife tests shows the GA has better global performance than the Baum–Welch algorithm. Results of jackknife tests show that this method performs better than all well known existing methods for topology predictions. Furthermore, it provides a function to predict SP in full-length TMBs sequences with fairish accuracy. Conclusion We show that our combined HMM-based method is a better choice for TMB topology prediction, which implements topology predictions with higher accuracy and additional SP predictions for full-length TMB sequences.

EEG-based estimation of mental fatigue by using KPCA–HMM and complexity parameters

Two complexity parameters of EEG, i.e. approximate entropy ( ApEn) and Kolmogorov complexity ( Kc) are utilized to characterize the complexity and irregularity of EEG data under the different mental fatigue states. Then the kernel principal component analysis (KPCA) and Hidden Markov Model (HMM) are combined to differentiate two mental fatigue states. The KPCA algorithm is employed to extract nonlinear features from the complexity parameters of EEG and improve the generalization performance of HMM. The investigation suggests that ApEn and Kc can effectively describe the dynamic complexity of EEG, which is strongly correlated with mental fatigue. Both complexity parameters are significantly decreased ( P < 0.005) as the mental fatigue level increases. These complexity parameters may be used as the indices of the mental fatigue level. Moreover, the joint KPCA–HMM method can effectively reduce the dimensionality of the feature vectors, accelerate the classification speed and achieve higher classification accuracy (84%) of mental fatigue. Hence KPCA–HMM could be a promising model for the estimation of mental fatigue.

Modeling action potential initiation using voltage-gated ion-channels kinetics from L5 pyramidal neurons of the rat cortex

Discussion • Sodium currents were measured from L5 pyramidal neurons using six stimulation protocols in the nucleated and outside-out configurations of the patch-clamp technique and fit to several kinetic models of voltage-gated sodium channels. • Out of a dozen voltage-gated sodium channel (VGSC) models (only relevant configurations shown) it appears that the most simplified, a four-state-16-parameter, Hidden-Markov-Model (HMM) provides the best fit for the data. • The ability of the chosen HMM model to simulate generation and propagation of action potentials (APs) in L5 pyramidal neurons was tested relative to the most prevalent and complete model of morphological influence on firing pattern. • The combination of the simplified HMM and a Hodgkin-Huxley-like voltage-gated potassium channel model provides the best simile of high frequency bursts of APs in L5 pyramidal neurons. • Further development and investigation of an adequate voltage-gated potassium channel HMM is needed to replace the HH paradigm. • The chosen HMM should be further tested on other types of VGSCs from various neurons and compared to the performance of other HMM. Figure 1: Fitting Newly Devised Models to Experimentally Recorded Voltage-Gated Na+ Channels Currents from neocortical pyramidal neurons –results of fitting a data set containing five pulse protocols to the kinetic model formalisms shown in the insets. A, A 16 kinetic parameters model consisting of two closed states, an open state and an inactivated state, with the maximal conductance of each protocol providing an extra parameter. The transition rate functions were defined as kij=aij.ezij .Vm and kji=aji.e-zji .Vm for i<j. B, An example of the fit achieved by the model shown in A to activation traces consisting of inward currents recorded in the nucleated configuration of the patch-clamp technique that were produced by a 100 ms pre-pulse to -130 mV followed by a 20 ms pulse to voltages between -130 to -25 mV at 5 mV increments. C, A 32 kinetic parameters model consisting of three closed states, an open state and two inactivated states, the maximal conductances and transition rate functions were included as in A. D, The same data as in B fit with the model shown in C. 1 ms 100 pA B

Augmented training of hidden Markov models to recognize remote homologs via simulated evolution

Motivation: While profile hidden Markov models (HMMs) are successful and powerful methods to recognize homologous proteins, they can break down when homology becomes too distant due to lack of sufficient training data. We show that we can improve the performance of HMMs in this domain by using a simple simulated model of evolution to create an augmented training set.Results: We show, in two different remote protein homolog tasks, that HMMs whose training is augmented with simulated evolution outperform HMMs trained only on real data. We find that a mutation rate between 15 and 20% performs best for recognizing G-protein coupled receptor proteins in different classes, and for recognizing SCOP super-family proteins from different families.Contacts: anoop.kumar@tufts.edu;lenore.cowen@tufts.edu

Classification of cries of infants with cleft-palate using parallel hidden Markov models

This paper addresses the problem of classification of infants with cleft palate. A hidden Markov model (HMM)-based cry classification algorithm is presented. A parallel HMM (PHMM) for coping with age masking, based on a maximum-likelihood decision rule, is introduced. The performance of the proposed algorithm under different model parameters and different feature sets is studied using a database of cries of infants with cleft palate (CLP). The proposed algorithm yields an average of 91% correct classification rate in a subject- and age-dependent experiment. In addition, it is shown that the PHMM significantly outperforms the HMM performance in classification of cries of CLP infants of different ages.

Classification of Online Game Players Using Action Transition Probability and Kullback Leibler Entropy

Online game players are more satisfied with contents tailored to their preferences. Player classification is necessary for determining which classes players belong to. In this paper, we propose a new player classification approach using action transition probability and Kullback Leibler entropy. In experiments with two online game simulators, Zereal and Simac, our approach performed better than an existing approach based on action frequency and comparably to another existing approach using the Hidden Markov Model (HMM). Our approach takes into account both the frequency and order of player action. While HMM performance depends on its structure and initial parameters, our approach requires no parameter settings.

Maximum Likelihood Linear Regression Adaptation for the Polynomial Segment Models

Speaker adaptation has long been applied to improve speech recognition performance of hidden Markov model (HMM)-based systems. Recently, the polynomial segment model (PSM) has been shown as a viable alternative that can significantly improve the performance of large vocabulary continuous speech recognition (LVCSR). In this letter, we extend the widely used HMM-based maximum likelihood linear regression (MLLR) speaker adaptation technique to PSMs. PSM properties, such as using segment as a modeling unit, and a polynomial curve as model mean, are taken into account in deriving the PSM-based MLLR. Experiments show that PSM-based MLLR adaptation performs equally well as the HMM-based MLLR adaptation with about 19% relative improvement from the SI model. In addition, another 5% relative improvement can be obtained by combining the adapted PSMs and HMMs

A comparative evaluation of neural networks and hidden Markov models for monitoring turning tool wear

Condition monitoring of machine tool inserts is important for increasing the reliability and quality of machining operations. Various methods have been proposed for effective tool condition monitoring (TCM), and currently it is generally accepted that the indirect sensor-based approach is the best practical solution to reliable TCM. Furthermore, in recent years, neural networks (NNs) have been shown to model successfully, the complex relationships between input feature sets of sensor signals and tool wear data. NNs have several properties that make them ideal for effectively handling noisy and even incomplete data sets. There are several NN paradigms which can be combined to model static and dynamic systems. Another powerful method of modeling noisy dynamic systems is by using hidden Markov models (HMMs), which are commonly employed in modern speech-recognition systems. The use of HMMs for TCM was recently proposed in the literature. Though the results of these studies were quite promising, no comparative results of competing methods such as NNs are currently available. This paper is aimed at presenting a comparative evaluation of the performance of NNs and HMMs for a TCM application. The methods are employed on exactly the same data sets obtained from an industrial turning operation. The advantages and disadvantages of both methods are described, which will assist the condition-monitoring community to choose a modeling method for other applications.

Improved profile HMM performance by assessment of critical algorithmic features in SAM and HMMER

BackgroundProfile hidden Markov model (HMM) techniques are among the most powerful methods for protein homology detection. Yet, the critical features for successful modelling are not fully known. In the present work we approached this by using two of the most popular HMM packages: SAM and HMMER. The programs' abilities to build models and score sequences were compared on a SCOP/Pfam based test set. The comparison was done separately for local and global HMM scoring.ResultsUsing default settings, SAM was overall more sensitive. SAM's model estimation was superior, while HMMER's model scoring was more accurate. Critical features for model building were then analysed by comparing the two packages' algorithmic choices and parameters. The weighting between prior probabilities and multiple alignment counts held the primary explanation why SAM's model building was superior. Our analysis suggests that HMMER gives too much weight to the sequence counts. SAM's emission prior probabilities were also shown to be more sensitive. The relative sequence weighting schemes are different in the two packages but performed equivalently.ConclusionSAM model estimation was more sensitive, while HMMER model scoring was more accurate. By combining the best algorithmic features from both packages the accuracy was substantially improved compared to their default performance.

Exploitation of unlabeled sequences in hidden Markov models

This paper presents a method for effectively using unlabeled sequential data in the learning of hidden Markov models (HMMs). With the conventional approach, class labels for unlabeled data are assigned deterministically by HMMs learned from labeled data. Such labeling often becomes unreliable when the number of labeled data is small. We propose an extended Baum-Welch (EBW) algorithm in which the labeling is undertaken probabilistically and iteratively so that the labeled and unlabeled data likelihoods are improved. Unlike the conventional approach, the EBW algorithm guarantees convergence to a local maximum of the likelihood. Experimental results on gesture data and speech data show that when labeled training data are scarce, by using unlabeled data, the EBW algorithm improves the classification performance of HMMs more robustly than the conventional naive labeling (NL) approach.

Use of unlabeled time series data in hidden Markov models

AbstractIn this paper, the authors propose a method of assigning probabilistic class labels to unlabeled time‐series data that can be collected inexpensively to use together with labeled data as training data in order to improve the performance of hidden Markov models (HMM) as classifiers. Conventional methods of using unlabeled data according to probabilistic class labels were only for static data. They could not deal with time‐series data, which are used with HMMs. Therefore, the authors extended the conventional methods as follows. First, they formed extended tied‐mixture hidden Markov models (ETM‐HMMs) by mixing multiple HMMs that exist independently for each class. Then, they introduced the extended Baum‐Welch (EBW) algorithm as the learning algorithm from time‐series data in which labeled and unlabeled data coexist in the ETM‐HMMs. The authors then assumed a situation in which there existed insufficient labeled training data and used the proposed method to perform classification experiments for Japanese sign language data and voice data. When unlabeled data were added to the labeled training data according to the proposed method, the classification error rate for Japanese sign language signs improved from 38.7% to 30.4%. Also, by making the model more detailed in addition to adding unlabeled data, the classification error rate for phonemes improved from 69.1% to 52.9%. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(13): 1–12, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10537

Non-parametric probability estimation for HMM-based automatic speech recognition

During the last decade, the most significant advances in the field of continuous speech recognition (CSR) have arisen from the use of hidden Markov models (HMM) for acoustic modeling. These models address one of the major issues for CSR: simultaneous modeling of temporal and frequency distortions in the speech signal. In the HMM, the temporal dimension is managed through an oriented states graph, each state accounting for the local frequency distortions through a probability density function. In this study, improvement of the HMM performance is expected from the introduction of a very effective non-parametric probability density function estimate: the k-nearest neighbors ( k-nn) estimate. First, experiments on a short-term speech spectrum identification task are performed to compare the k-nn estimate and the widespread estimate based on mixtures of Gaussian functions. Then adaptations implied by the integration of the k-nn estimate in an HMM-based recognition system are developed. An optimal training protocol is obtained based on the introduction of the membership coefficients in the HMM parameters. The membership coefficients measure the degree of association between a reference acoustic vector and a HMM state. The training procedure uses the expectation-maximization (EM) algorithm applied to the membership coefficient estimation. Its convergence is shown according to the maximum likelihood criterion. This study leads to the development of a baseline k-nn/HMM recognition system which is evaluated on the TIMIT speech database. Further improvements of the k-nn/HMM system are finally sought through the introduction of a temporal information into the representation space ( delta coefficients) and the adaptation of the references (mainly, gender modeling and contextual modeling).

A segmental-feature HMM for continuous speech recognition based on a parametric trajectory model

In this paper, we propose a new acoustic model for characterizing segmental features and an algorithm based upon a general framework of hidden Markov models (HMMs). The segmental features are represented as a trajectory of observed vector sequences by a polynomial regression function. To obtain the polynomial trajectory from speech segments, we modify the design matrix to include transitional information for contiguous frames. We also propose methods for estimating the likelihood of a given segment and trajectory parameters. The observation probability of a given segment is represented as the relation between the segment likelihood and the estimation error of the trajectories. The estimation error of a trajectory is considered the weight of the likelihood of a given segment in a state. This weight represents the probability of how well the corresponding trajectory characterizes the segment. The proposed model can be regarded as a generalization of a conventional HMM and a parametric trajectory model. We conducted several experiments to establish the effectiveness of the proposed method and the characteristics of the segmental features. The recognition results on the TIMIT database demonstrate that the performance of segmental-feature HMM (SFHMM) is better than that of a conventional HMM.

Assignment of homology to genome sequences using a library of hidden Markov models that represent all proteins of known structure

Of the sequence comparison methods, profile-based methods perform with greater selectively than those that use pairwise comparisons. Of the profile methods, hidden Markov models (HMMs) are apparently the best. The first part of this paper describes calculations that (i) improve the performance of HMMs and (ii) determine a good procedure for creating HMMs for sequences of proteins of known structure. For a family of related proteins, more homologues are detected using multiple models built from diverse single seed sequences than from one model built from a good alignment of those sequences. A new procedure is described for detecting and correcting those errors that arise at the model-building stage of the procedure. These two improvements greatly increase selectivity and coverage. The second part of the paper describes the construction of a library of HMMs, called SUPERFAMILY, that represent essentially all proteins of known structure. The sequences of the domains in proteins of known structure, that have identities less than 95 %, are used as seeds to build the models. Using the current data, this gives a library with 4894 models. The third part of the paper describes the use of the SUPERFAMILY model library to annotate the sequences of over 50 genomes. The models match twice as many target sequences as are matched by pairwise sequence comparison methods. For each genome, close to half of the sequences are matched in all or in part and, overall, the matches cover 35 % of eukaryotic genomes and 45 % of bacterial genomes. On average roughly 15% of genome sequences are labelled as being hypothetical yet homologous to proteins of known structure. The annotations derived from these matches are available from a public web server at: http://stash.mrc-lmb.cam.ac.uk/SUPERFAMILY. This server also enables users to match their own sequences against the SUPERFAMILY model library.

Noise-compensated hidden Markov models

The technique of hidden Markov models has been established as one of the most successful methods applied to the problem of speech recognition. However, its performance is considerably degraded when the speech signal is contaminated by noise. This work presents a technique which improves the performance of hidden Markov models when these models are used in different noise conditions during the speech recognition process. The input speech signal enters unchanged to the recognition process, while the models used by the recognition system are compensated according to the affecting noise characteristics, power and spectral shape. Hence, the compensation stage is independent of the recognition stage, allowing the models to be continually adjusted. The models used in this work are from a continuous density hidden Markov algorithm, having cepstral coefficients derived from linear predictive analysis as state parameters. It is used only static features in the models in order to show that, when properly compensated for the noise, these static features contribute significantly to improve noisy speech recognition. It is observed from the results that the parameters kept their capability to discriminate among different classes of signals, indicating that, in the context of speech recognition, the use of autoregressive-derived parameters with noisy signals does not represent an impediment. A matrix-way of converting from autoregressive coefficients to normalized autocorrelation coefficients is presented. The affecting noise is assumed additive and statistically independent of the speech signal. Although the noise dealt with should also be stationary, good performance was achieved for nonstationary noise, such as operations room noise and factory environment noise. The concept of intra-word signal-to-noise ratio is presented and successfully applied. The resulting compensated models are revealed to be less dependent on the training data set when compared to the trained hidden Markov models. Due to the computational simplicity, the time required to adjust a model is significantly shorter than the time to train it.

A hybrid segmental neural net/hidden Markov model system for continuous speech recognition

The current state-of-the-art in large-vocabulary, continuous speech recognition is based on the use of hidden Markov models (HMM). In an attempt to improve over HMM performance, the authors developed a hybrid system that combines the advantages of neural networks and HMM using a multiple hypothesis (or N-best) paradigm. The connectionist component of the system, the segmental neural net (SNN), models all the frames of a phonetic segment simultaneously, thus overcoming the well-known conditional-independence limitation of the HMM. They describe the hybrid system and discuss various aspects of SNN modeling, including network architectures, training algorithms and context modeling. Finally, they evaluate the hybrid system by performing several speaker-independent experiments with the DARPA Resource Management (RM) corpus, and demonstrate that the hybrid system shows a consistent improvement in performance over the baseline HMM system. >

Continuous hidden Markov models integrating transitional and instantaneous features for Mandarin syllable recognition

Feature parameters describing spectral transitions of speech signals have been properly integrated with the instantaneous features in many different approaches proposed for speech recognition, and significant performance improvements have been attained. Most of these methods are designed for recognition systems based on dynamic time warping (DTW) or discrete hidden Markov models (HMM). However, it has been experimentally shown that for the difficult problem of recognizing the highly confusing Mandarin syllables with limited amount of training data, the performances of DTW and discrete HMM techniques are much worse than that of continuous HMMs. In this paper, the performance of continuous HMMs using one type of transitional features in speaker-dependent recognition of the highly confusing Mandarin syllables is first evaluated and discussed in detail under the constraint of very limited training data. Three approaches are then proposed to integrate the instantaneous and transitional features for recognition systems based on continuous hidden Markov models. They are the most straightforward concatenation—integration approach in which the instantaneous and transitional feature vectors are simply concatenated, the two-maximization approach in which the output distribution functions for the instantaneous and transitional feature vectors are maximized separately, and the two-model approach in which two HMMs respectively for instantaneous and transitional feature vectors are independently trained but the log likelihoods are summed up with proper weighting. After extensive experiments and careful analysis, it is found that the three approaches respectively provide attractive performance under different conditions. For example, with the two-maximization approach a recognition rate (93·89%) only slightly lower than the highest achievable rate for the concatenation-integration approach (94·36% for M = 5) can be obtained at a much smaller number of mixtures ( M = 2).