Semi-supervised learning of statistical models for natural language understanding

Natural language understanding is to specify a computational model that maps sentences to their semantic mean representation. In this paper, we propose a novel framework to train the statistical models without using expensive fully annotated data. In particular, the input of our framework is a set of sentences labeled with abstract semantic annotations. These annotations encode the underlying embedded semantic structural relations without explicit word/semantic tag alignment. The proposed framework can automatically induce derivation rules that map sentences to their semantic meaning representations. The learning framework is applied on two statistical models, the conditional random fields (CRFs) and the hidden Markov support vector machines (HM-SVMs). Our experimental results on the DARPA communicator data show that both CRFs and HM-SVMs outperform the baseline approach, previously proposed hidden vector state (HVS) model which is also trained on abstract semantic annotations. In addition, the proposed framework shows superior performance than two other baseline approaches, a hybrid framework combining HVS and HM-SVMs and discriminative training of HVS, with a relative error reduction rate of about 25% and 15% being achieved in F-measure.

Modelling financial time series with switching state space models

The deficiencies of stationary models applied to financial time series are well documented. A special form of non-stationarity, where the underlying generator switches between (approximately) stationary regimes, seems particularly appropriate for financial markets. We use a dynamic switching (modelled by a hidden Markov model) combined with a linear dynamical system in a hybrid switching state space model (SSSM) and discuss the practical details of training such models with a variational EM algorithm due to [Ghahramani and Hilton,1998]. The performance of the SSSM is evaluated on several financial data sets and it is shown to improve on a number of existing benchmark methods.

Dynamical local models for segmentation and prediction of financial time series

In the analysis and prediction of many real-world time series, the assumption of stationarity is not valid. A special form of non-stationarity, where the underlying generator switches between (approximately) stationary regimes, seems particularly appropriate for financial markets. We introduce a new model which combines a dynamic switching (controlled by a hidden Markov model) and a non-linear dynamical system. We show how to train this hybrid model in a maximum likelihood approach and evaluate its performance on both synthetic and financial data.

Discriminative training of the hidden vector state model for semantic parsing

In this paper, we discuss how discriminative training can be applied to the hidden vector state (HVS) model in different task domains. The HVS model is a discrete hidden Markov model (HMM) in which each HMM state represents the state of a push-down automaton with a finite stack size. In previous applications, maximum-likelihood estimation (MLE) is used to derive the parameters of the HVS model. However, MLE makes a number of assumptions and unfortunately some of these assumptions do not hold. Discriminative training, without making such assumptions, can improve the performance of the HVS model by discriminating the correct hypothesis from the competing hypotheses. Experiments have been conducted in two domains: the travel domain for the semantic parsing task using the DARPA Communicator data and the Air Travel Information Services (ATIS) data and the bioinformatics domain for the information extraction task using the GENIA corpus. The results demonstrate modest improvements of the performance of the HVS model using discriminative training. In the travel domain, discriminative training of the HVS model gives a relative error reduction rate of 31 percent in F-measure when compared with MLE on the DARPA Communicator data and 9 percent on the ATIS data. In the bioinformatics domain, a relative error reduction rate of 4 percent in F-measure is achieved on the GENIA corpus.

Toward the prediction of class I and II mouse major histocompatibility complex-peptide-binding affinity:in silico bioinformatic step-by-step guide using quantitative structure-activity relationships

Quantitative structure-activity relationship (QSAR) analysis is a cornerstone of modern informatics. Predictive computational models of peptide-major histocompatibility complex (MHC)-binding affinity based on QSAR technology have now become important components of modern computational immunovaccinology. Historically, such approaches have been built around semiqualitative, classification methods, but these are now giving way to quantitative regression methods. We review three methods--a 2D-QSAR additive-partial least squares (PLS) and a 3D-QSAR comparative molecular similarity index analysis (CoMSIA) method--which can identify the sequence dependence of peptide-binding specificity for various class I MHC alleles from the reported binding affinities (IC50) of peptide sets. The third method is an iterative self-consistent (ISC) PLS-based additive method, which is a recently developed extension to the additive method for the affinity prediction of class II peptides. The QSAR methods presented here have established themselves as immunoinformatic techniques complementary to existing methodology, useful in the quantitative prediction of binding affinity: current methods for the in silico identification of T-cell epitopes (which form the basis of many vaccines, diagnostics, and reagents) rely on the accurate computational prediction of peptide-MHC affinity. We have reviewed various human and mouse class I and class II allele models. Studied alleles comprise HLA-A*0101, HLA-A*0201, HLA-A*0202, HLA-A*0203, HLA-A*0206, HLA-A*0301, HLA-A*1101, HLA-A*3101, HLA-A*6801, HLA-A*6802, HLA-B*3501, H2-K(k), H2-K(b), H2-D(b) HLA-DRB1*0101, HLA-DRB1*0401, HLA-DRB1*0701, I-A(b), I-A(d), I-A(k), I-A(S), I-E(d), and I-E(k). In this chapter we show a step-by-step guide into predicting the reliability and the resulting models to represent an advance on existing methods. The peptides used in this study are available from the AntiJen database (http://www.jenner.ac.uk/AntiJen). The PLS method is available commercially in the SYBYL molecular modeling software package. The resulting models, which can be used for accurate T-cell epitope prediction, will be made are freely available online at the URL http://www.jenner.ac.uk/MHCPred.

Class I T-cell epitope prediction:improvements using a combination of proteasome cleavage, TAP affinity, and MHC binding

Cleavage by the proteasome is responsible for generating the C terminus of T-cell epitopes. Modeling the process of proteasome cleavage as part of a multi-step algorithm for T-cell epitope prediction will reduce the number of non-binders and increase the overall accuracy of the predictive algorithm. Quantitative matrix-based models for prediction of the proteasome cleavage sites in a protein were developed using a training set of 489 naturally processed T-cell epitopes (nonamer peptides) associated with HLA-A and HLA-B molecules. The models were validated using an external test set of 227 T-cell epitopes. The performance of the models was good, identifying 76% of the C-termini correctly. The best model of proteasome cleavage was incorporated as the first step in a three-step algorithm for T-cell epitope prediction, where subsequent steps predicted TAP affinity and MHC binding using previously derived models.

Toward prediction of class II mouse major histocompatibility complex peptide binding affinity:in silico bioinformatic evaluation using partial least squares, a robust multivariate statistical technique

The accurate identification of T-cell epitopes remains a principal goal of bioinformatics within immunology. As the immunogenicity of peptide epitopes is dependent on their binding to major histocompatibility complex (MHC) molecules, the prediction of binding affinity is a prerequisite to the reliable prediction of epitopes. The iterative self-consistent (ISC) partial-least-squares (PLS)-based additive method is a recently developed bioinformatic approach for predicting class II peptide−MHC binding affinity. The ISC−PLS method overcomes many of the conceptual difficulties inherent in the prediction of class II peptide−MHC affinity, such as the binding of a mixed population of peptide lengths due to the open-ended class II binding site. The method has applications in both the accurate prediction of class II epitopes and the manipulation of affinity for heteroclitic and competitor peptides. The method is applied here to six class II mouse alleles (I-Ab, I-Ad, I-Ak, I-As, I-Ed, and I-Ek) and included peptides up to 25 amino acids in length. A series of regression equations highlighting the quantitative contributions of individual amino acids at each peptide position was established. The initial model for each allele exhibited only moderate predictivity. Once the set of selected peptide subsequences had converged, the final models exhibited a satisfactory predictive power. Convergence was reached between the 4th and 17th iterations, and the leave-one-out cross-validation statistical terms - q2, SEP, and NC - ranged between 0.732 and 0.925, 0.418 and 0.816, and 1 and 6, respectively. The non-cross-validated statistical terms r2 and SEE ranged between 0.98 and 0.995 and 0.089 and 0.180, respectively. The peptides used in this study are available from the AntiJen database (http://www.jenner.ac.uk/AntiJen). The PLS method is available commercially in the SYBYL molecular modeling software package. The resulting models, which can be used for accurate T-cell epitope prediction, will be made freely available online (http://www.jenner.ac.uk/MHCPred).

GTM through time

The standard GTM (generative topographic mapping) algorithm assumes that the data on which it is trained consists of independent, identically distributed (iid) vectors. For time series, however, the iid assumption is a poor approximation. In this paper we show how the GTM algorithm can be extended to model time series by incorporating it as the emission density in a hidden Markov model. Since GTM has discrete hidden states we are able to find a tractable EM algorithm, based on the forward-backward algorithm, to train the model. We illustrate the performance of GTM through time using flight recorder data from a helicopter.

Combining spatially distributed predictions from neural networks

In this report we discuss the problem of combining spatially-distributed predictions from neural networks. An example of this problem is the prediction of a wind vector-field from remote-sensing data by combining bottom-up predictions (wind vector predictions on a pixel-by-pixel basis) with prior knowledge about wind-field configurations. This task can be achieved using the scaled-likelihood method, which has been used by Morgan and Bourlard (1995) and Smyth (1994), in the context of Hidden Markov modelling

A generative probabilistic oriented wavelet model for texture segmentation

This Letter addresses image segmentation via a generative model approach. A Bayesian network (BNT) in the space of dyadic wavelet transform coefficients is introduced to model texture images. The model is similar to a Hidden Markov model (HMM), but with non-stationary transitive conditional probability distributions. It is composed of discrete hidden variables and observable Gaussian outputs for wavelet coefficients. In particular, the Gabor wavelet transform is considered. The introduced model is compared with the simplest joint Gaussian probabilistic model for Gabor wavelet coefficients for several textures from the Brodatz album [1]. The comparison is based on cross-validation and includes probabilistic model ensembles instead of single models. In addition, the robustness of the models to cope with additive Gaussian noise is investigated. We further study the feasibility of the introduced generative model for image segmentation in the novelty detection framework [2]. Two examples are considered: (i) sea surface pollution detection from intensity images and (ii) image segmentation of the still images with varying illumination across the scene.

Towards sea surface pollution detection from visible band images

This paper presents a novel approach to water pollution detection from remotely sensed low-platform mounted visible band camera images. We examine the feasibility of unsupervised segmentation for slick (oily spills on the water surface) region labelling. Adaptive and non adaptive filtering is combined with density modeling of the obtained textural features. A particular effort is concentrated on the textural feature extraction from raw intensity images using filter banks and adaptive feature extraction from the obtained output coefficients. Segmentation in the extracted feature space is achieved using Gaussian mixture models (GMM).

Volatility changes caused by the trading system:a Markov switching application

An expanding literature exists to suggest that the trading mechanism can influence the volatility of security returns. This study adds to this literature by examining the impact that the introduction of SETS, on the London Stock Exchange, had on the volatility of security returns. Using a Markov switching regime change model security volatility is categorized as being in a regime of either high or low volatility. It is shown that prior to the introduction of SETS securities tended to be in a low volatility regime. At the time SETS was introduced securities moved to a high volatility regime. This suggests that volatility increased when SETS was introduced.

DEA models for ratio data:convexity consideration

Data envelopment analysis (DEA) is defined based on observed units and by finding the distance of each unit to the border of estimated production possibility set (PPS). The convexity is one of the underlying assumptions of the PPS. This paper shows some difficulties of using standard DEA models in the presence of input-ratios and/or output-ratios. The paper defines a new convexity assumption when data includes a ratio variable. Then it proposes a series of modified DEA models which are capable to rectify this problem.

Identifying and correcting non-Markov states in peptide conformational dynamics

Conformational transitions in proteins define their biological activity and can be investigated in detail using the Markov state model. The fundamental assumption on the transitions between the states, their Markov property, is critical in this framework. We test this assumption by analyzing the transitions obtained directly from the dynamics of a molecular dynamics simulated peptide valine-proline-alanine-leucine and states defined phenomenologically using clustering in dihedral space. We find that the transitions are Markovian at the time scale of ˜ 50 ps and longer. However, at the time scale of 30–40 ps the dynamics loses its Markov property. Our methodology reveals the mechanism that leads to non-Markov behavior. It also provides a way of regrouping the conformations into new states that now possess the required Markov property of their dynamics.

Neural Network Modelling with Input Uncertainty: Theory and Application

It is generally assumed when using Bayesian inference methods for neural networks that the input data contains no noise. For real-world (errors in variable) problems this is clearly an unsafe assumption. This paper presents a Bayesian neural network framework which accounts for input noise provided that a model of the noise process exists. In the limit where the noise process is small and symmetric it is shown, using the Laplace approximation, that this method adds an extra term to the usual Bayesian error bar which depends on the variance of the input noise process. Further, by treating the true (noiseless) input as a hidden variable, and sampling this jointly with the network’s weights, using a Markov chain Monte Carlo method, it is demonstrated that it is possible to infer the regression over the noiseless input. This leads to the possibility of training an accurate model of a system using less accurate, or more uncertain, data. This is demonstrated on both the, synthetic, noisy sine wave problem and a real problem of inferring the forward model for a satellite radar backscatter system used to predict sea surface wind vectors.