Speaker Recognition In Noisy Conditions With Limited Training Data

In this paper we present a novel method for performing speaker recognition with very limited training data and in the presence of background noise. Similarity-based speaker recognition is considered so that speaker models can be created with limited training speech data. The proposed similarity is a form of cosine similarity used as a distance measure between speech feature vectors. Each speech frame is modelled using subband features, and into this framework, multicondition training and optimal feature selection are introduced, making the system capable of performing speaker recognition in the presence of realistic, time-varying noise, which is unknown during training. Speaker identi?cation experiments were carried out using the SPIDRE database. The performance of the proposed new system for noise compensation is compared to that of an oracle model; the speaker identi?cation accuracy for clean speech by the new system trained with limited training data is compared to that of a GMM trained with several minutes of speech. Both comparisons have demonstrated the effectiveness of the new model. Finally, experiments were carried out to test the new model for speaker identi?cation given limited training data and with differing levels and types of realistic background noise. The results have demonstrated the robustness of the new system.

Collectives of diagnostic biomarkers identify high-risk subpopulations of hematuria patients: exploiting heterogeneity in large-scale biomarker data

Background: Ineffective risk stratification can delay diagnosis of serious disease in patients with hematuria. We applied a systems biology approach to analyze clinical, demographic and biomarker measurements (n = 29) collected from 157 hematuric patients: 80 urothelial cancer (UC) and 77 controls with confounding pathologies.

Methods: On the basis of biomarkers, we conducted agglomerative hierarchical clustering to identify patient and biomarker clusters. We then explored the relationship between the patient clusters and clinical characteristics using Chi-square analyses. We determined classification errors and areas under the receiver operating curve of Random Forest Classifiers (RFC) for patient subpopulations using the biomarker clusters to reduce the dimensionality of the data.

Results: Agglomerative clustering identified five patient clusters and seven biomarker clusters. Final diagnoses categories were non-randomly distributed across the five patient clusters. In addition, two of the patient clusters were enriched with patients with ‘low cancer-risk’ characteristics. The biomarkers which contributed to the diagnostic classifiers for these two patient clusters were similar. In contrast, three of the patient clusters were significantly enriched with patients harboring ‘high cancer-risk” characteristics including proteinuria, aggressive pathological stage and grade, and malignant cytology. Patients in these three clusters included controls, that is, patients with other serious disease and patients with cancers other than UC. Biomarkers which contributed to the diagnostic classifiers for the largest ‘high cancer- risk’ cluster were different than those contributing to the classifiers for the ‘low cancer-risk’ clusters. Biomarkers which contributed to subpopulations that were split according to smoking status, gender and medication were different.

Conclusions: The systems biology approach applied in this study allowed the hematuric patients to cluster naturally on the basis of the heterogeneity within their biomarker data, into five distinct risk subpopulations. Our findings highlight an approach with the promise to unlock the potential of biomarkers. This will be especially valuable in the field of diagnostic bladder cancer where biomarkers are urgently required. Clinicians could interpret risk classification scores in the context of clinical parameters at the time of triage. This could reduce cystoscopies and enable priority diagnosis of aggressive diseases, leading to improved patient outcomes at reduced costs. © 2013 Emmert-Streib et al; licensee BioMed Central Ltd.

An Investigation Into Features For Multi-View Lipreading

For the first time in this paper we present results showing the effect of speaker head pose angle on automatic lip-reading performance over a wide range of closely spaced angles. We analyse the effect head pose has upon the features themselves and show that by selecting coefficients with minimum variance w.r.t. pose angle, recognition performance can be improved when train-test pose angles differ. Experiments are conducted using the initial phase of a unique multi view Audio-Visual database designed specifically for research and development of pose-invariant lip-reading systems. We firstly show that it is the higher order horizontal spatial frequency components that become most detrimental as the pose deviates. Secondly we assess the performance of different feature selection masks across a range of pose angles including a new mask based on Minimum Cross-Pose Variance coefficients. We report a relative improvement of 50% in Word Error Rate when using our selection mask over a common energy based selection during profile view lip-reading.

Robust Multimodal Person Identification With Limited Training Data

This paper presents a novel method of audio-visual feature-level fusion for person identification where both the speech and facial modalities may be corrupted, and there is a lack of prior knowledge about the corruption. Furthermore, we assume there are limited amount of training data for each modality (e.g., a short training speech segment and a single training facial image for each person). A new multimodal feature representation and a modified cosine similarity are introduced to combine and compare bimodal features with limited training data, as well as vastly differing data rates and feature sizes. Optimal feature selection and multicondition training are used to reduce the mismatch between training and testing, thereby making the system robust to unknown bimodal corruption. Experiments have been carried out on a bimodal dataset created from the SPIDRE speaker recognition database and AR face recognition database with variable noise corruption of speech and occlusion in the face images. The system's speaker identification performance on the SPIDRE database, and facial identification performance on the AR database, is comparable with the literature. Combining both modalities using the new method of multimodal fusion leads to significantly improved accuracy over the unimodal systems, even when both modalities have been corrupted. The new method also shows improved identification accuracy compared with the bimodal systems based on multicondition model training or missing-feature decoding alone.

JPEG Image Steganalysis Using Multivariate PDF Estimates With MRF Cliques

Blind steganalysis of JPEG images is addressed by modeling the correlations among the DCT coefficients using K -variate (K = 2) p.d.f. estimates (p.d.f.s) constructed by means of Markov random field (MRF) cliques. The reasoning of using high variate p.d.f.s together with MRF cliques for image steganalysis is explained via a classical detection problem. Although our approach has many improvements over the current state-of-the-art, it suffers from the high dimensionality and the sparseness of the high variate p.d.f.s. The dimensionality problem as well as the sparseness problem are solved heuristically by means of dimensionality reduction and feature selection algorithms. The detection accuracy of the proposed method(s) is evaluated over Memon's (30.000 images) and Goljan's (1912 images) image sets. It is shown that practically applicable steganalysis systems are possible with a suitable dimensionality reduction technique and these systems can provide, in general, improved detection accuracy over the current state-of-the-art. Experimental results also justify this assertion.

Decoding semantics across fMRI sessions with different stimulus modalities:a practical MVPA study

Both embodied and symbolic accounts of conceptual organization would predict partial sharing and partial differentiation between the neural activations seen for concepts activated via different stimulus modalities. But cross-participant and cross-session variability in BOLD activity patterns makes analyses of such patterns with MVPA methods challenging. Here, we examine the effect of cross-modal and individual variation on the machine learning analysis of fMRI data recorded during a word property generation task. We present the same set of living and non-living concepts (land-mammals, or work tools) to a cohort of Japanese participants in two sessions: the first using auditory presentation of spoken words; the second using visual presentation of words written in Japanese characters. Classification accuracies confirmed that these semantic categories could be detected in single trials, with within-session predictive accuracies of 80-90%. However cross-session prediction (learning from auditory-task data to classify data from the written-word-task, or vice versa) suffered from a performance penalty, achieving 65-75% (still individually significant at p « 0.05). We carried out several follow-on analyses to investigate the reason for this shortfall, concluding that distributional differences in neither time nor space alone could account for it. Rather, combined spatio-temporal patterns of activity need to be identified for successful cross-session learning, and this suggests that feature selection strategies could be modified to take advantage of this.

Root Cause Analysis by a Combined Sparse Classification and Monte Carlo Approach

Classification methods with embedded feature selection capability are very appealing for the analysis of complex processes since they allow the analysis of root causes even when the number of input variables is high. In this work, we investigate the performance of three techniques for classification within a Monte Carlo strategy with the aim of root cause analysis. We consider the naive bayes classifier and the logistic regression model with two different implementations for controlling model complexity, namely, a LASSO-like implementation with a L1 norm regularization and a fully Bayesian implementation of the logistic model, the so called relevance vector machine. Several challenges can arise when estimating such models mainly linked to the characteristics of the data: a large number of input variables, high correlation among subsets of variables, the situation where the number of variables is higher than the number of available data points and the case of unbalanced datasets. Using an ecological and a semiconductor manufacturing dataset, we show advantages and drawbacks of each method, highlighting the superior performance in term of classification accuracy for the relevance vector machine with respect to the other classifiers. Moreover, we show how the combination of the proposed techniques and the Monte Carlo approach can be used to get more robust insights into the problem under analysis when faced with challenging modelling conditions.

Using Brain Data for Sentiment Analysis

We present the results of exploratory experiments using lexical valence extracted from brain using electroencephalography (EEG) for sentiment analysis. We selected 78 English words (36 for training and 42 for testing), presented as stimuli to 3 English native speakers. EEG signals were recorded from the subjects while they performed a mental imaging task for each word stimulus. Wavelet decomposition was employed to extract EEG features from the time-frequency domain. The extracted features were used as inputs to a sparse multinomial logistic regression (SMLR) classifier for valence classification, after univariate ANOVA feature selection. After mapping EEG signals to sentiment valences, we exploited the lexical polarity extracted from brain data for the prediction of the valence of 12 sentences taken from the SemEval-2007 shared task, and compared it against existing lexical resources.

Robust bimodal person identification using face and speech with limited training data and corruption of both modalities

This paper presents a novel method of audio-visual fusion for person identification where both the speech and facial modalities may be corrupted, and there is a lack of prior knowledge about the corruption. Furthermore, we assume there is a limited amount of training data for each modality (e.g., a short training speech segment and a single training facial image for each person). A new representation and a modified cosine similarity are introduced for combining and comparing bimodal features with limited training data as well as vastly differing data rates and feature sizes. Optimal feature selection and multicondition training are used to reduce the mismatch between training and testing, thereby making the system robust to unknown bimodal corruption. Experiments have been carried out on a bimodal data set created from the SPIDRE and AR databases with variable noise corruption of speech and occlusion in the face images. The new method has demonstrated improved recognition accuracy.

Network signatures based on gene pair expression ratios improve classification and the analysis of muscle-invasive urothelial cancer

Urothelial cancer (UC) is highly recurrent and can progress from non-invasive (NMIUC) to a more aggressive muscle-invasive (MIUC) subtype that invades the muscle tissue layer of the bladder. We present a proof of principle study that network-based features of gene pairs can be used to improve classifier performance and the functional analysis of urothelial cancer gene expression data. In the first step of our procedure each individual sample of a UC gene expression dataset is inflated by gene pair expression ratios that are defined based on a given network structure. In the second step an elastic net feature selection procedure for network-based signatures is applied to discriminate between NMIUC and MIUC samples. We performed a repeated random subsampling cross validation in three independent datasets. The network signatures were characterized by a functional enrichment analysis and studied for the enrichment of known cancer genes. We observed that the network-based gene signatures from meta collections of proteinprotein interaction (PPI) databases such as CPDB and the PPI databases HPRD and BioGrid improved the classification performance compared to single gene based signatures. The network based signatures that were derived from PPI databases showed a prominent enrichment of cancer genes (e.g., TP53, TRIM27 and HNRNPA2Bl). We provide a novel integrative approach for large-scale gene expression analysis for the identification and development of novel diagnostical targets in bladder cancer. Further, our method allowed to link cancer gene associations to network-based expression signatures that are not observed in gene-based expression signatures.

Machine learning methods for quantitative structure-property relationship modeling

Tese de doutoramento, Informática (Bioinformática), Universidade de Lisboa, Faculdade de Ciências, 2014

Detecting Adverse Events in Clinical Trial Free Text

Thesis (Master's)--University of Washington, 2013

Inferring Big 5 Personality from Online Social Networks

Thesis (Master's)--University of Washington, 2014

Risk Modelling Framework for Emergency Hospital Readmission, Using Hospital Episode Statistics Inpatient Data

The objective of this study was to develop, test and benchmark a framework and a predictive risk model for hospital emergency readmission within 12 months. We performed the development using routinely collected Hospital Episode Statistics data covering inpatient hospital admissions in England. Three different timeframes were used for training, testing and benchmarking: 1999 to 2004, 2000 to 2005 and 2004 to 2009 financial years. Each timeframe includes 20% of all inpatients admitted within the trigger year. The comparisons were made using positive predictive value, sensitivity and specificity for different risk cut-offs, risk bands and top risk segments, together with the receiver operating characteristic curve. The constructed Bayes Point Machine using this feature selection framework produces a risk probability for each admitted patient, and it was validated for different timeframes, sub-populations and cut-off points. At risk cut-off of 50%, the positive predictive value was 69.3% to 73.7%, the specificity was 88.0% to 88.9% and sensitivity was 44.5% to 46.3% across different timeframes. Also, the area under the receiver operating characteristic curve was 73.0% to 74.3%. The developed framework and model performed considerably better than existing modelling approaches with high precision and moderate sensitivity.

Statistical classification of road pavements using near field vehicle rolling noise measurements

Low noise surfaces have been increasingly considered as a viable and cost-effective alternative to acoustical barriers. However, road planners and administrators frequently lack information on the correlation between the type of road surface and the resulting noise emission profile. To address this problem, a method to identify and classify different types of road pavements was developed, whereby near field road noise is analyzed using statistical learning methods. The vehicle rolling sound signal near the tires and close to the road surface was acquired by two microphones in a special arrangement which implements the Close-Proximity method. A set of features, characterizing the properties of the road pavement, was extracted from the corresponding sound profiles. A feature selection method was used to automatically select those that are most relevant in predicting the type of pavement, while reducing the computational cost. A set of different types of road pavement segments were tested and the performance of the classifier was evaluated. Results of pavement classification performed during a road journey are presented on a map, together with geographical data. This procedure leads to a considerable improvement in the quality of road pavement noise data, thereby increasing the accuracy of road traffic noise prediction models.