We consider an online learning scenario in which the learner can make predictions on the basis of a fixed set of experts. We derive upper and lower relative loss bounds for a class of universal learning algorithms involving a switching dynamics over the choice of the experts. On the basis of the performance bounds we provide the optimal a priori discretization of the switching-rate parameter that governs the switching dynamics. We demonstrate the algorithm in the context of wireless networks.

Dissertação (mestrado)—Universidade de Brasilia, Faculdade de Tecnologia, Departamento de Engenharia Elétrica, Programa de Pós-Graduação em Engenharia Elétrica, 2011. / Submitted by Alaíde Gonçalves dos Santos (alaide@unb.br) on 2012-09-18T14:00:45Z No. of bitstreams: 1 2011_OsvaldoDalbenJunior.pdf: 3404533 bytes, checksum: 553bcf63bc1e8bcf2d4970f112a97520 (MD5) / Approved for entry into archive by Leandro Silva Borges(leandroborges@bce.unb.br) on 2012-09-19T21:17:42Z (GMT) No. of bitstreams: 1 2011_OsvaldoDalbenJunior.pdf: 3404533 bytes, checksum: 553bcf63bc1e8bcf2d4970f112a97520 (MD5) / Made available in DSpace on 2012-09-19T21:17:42Z (GMT). No. of bitstreams: 1 2011_OsvaldoDalbenJunior.pdf: 3404533 bytes, checksum: 553bcf63bc1e8bcf2d4970f112a97520 (MD5) / Um dos atuais desafios na área da forense computacional está relacionado à análise de mídias computacionais apreendidas em grande quantidade pelas orças policiais. Os arquivos armazenados nessas mídias podem conter nomes de pessoas e organizações suspeitos, porém desconhecidos pelas equipes de investigação. O presente trabalho propõe a criação de um modelo de Reconhecimento de Entidades Mencionadas (REM) baseado no Modelo Oculto de Markov (HMM) para extrair nomes de pessoas e organizações de textos não estruturados contidos em mídias apreendidas. O modelo proposto, denominado ICCHMM {Identiication-Classification Context HMM é dividido em dois submodelos - identificação e classificação de entidades - e utiliza as informações do contexto das palavras e um gazetteer como forma de obter melhor desempenho. Experimentos foram realizados aplicados a corpora públicos e forenses e os resultados do ICC-HMM superaram os obtidos por sistemas participantes de avaliações conjuntas específicas para o REM no idioma português, o que sugere que o modelo proposto é aplicável ao cenário forense nacional. _______________________________________________________________________________________ ABSTRACT / One of the current challenges in computer forensics is related to the analysis of computer media seized in large quantities by he police. Files stored in these media may contain names of people and organizations suspected, but unknown by the analysis teams. This paper proposes the creation of a named entity recognition (NER) model based on the Hidden Markov Model (HMM) to extract names of people and organizations contained in unstructured text of seized media. The proposed model, called ICC-HMM (Identification - Classification Context HMM) is divided into two sub-models - identification and classification of entities - and uses the context information of words and a gazetteer in order to obtain better performance. Experiments were carried out on forensic corpora and our results outperformed some of the best NER-based systems in Portuguese language. This suggests that the proposed model is applicable in brazilian computer forensics

Made available in DSpace on 2014-06-12T15:58:49Z (GMT). No. of bitstreams: 2 arquivo4727_1.pdf: 1069518 bytes, checksum: f7d66b889164069c63fdfafe11edf163 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2004 / Existe atualmente uma gigantesca quantidade de informações disponibilizada em formato de texto na Internet e nas redes das grandes corporações. Essas informações encontram-se em um formato não estruturado, dificilmente manipulável por programas de computador. A Extração de Informação (EI) tem como objetivo transformar documentos textuais em um formato estruturado, mapeando a informação contida em um documento em uma estrutura tabular. Tal estrutura é mais facilmente tratável por programas de computador, possibilitando assim a sua utilização por variadas aplicações inteligentes . Dentro da Inteligência Artificial, vemos duas abordagens para tratar o problema da EI: os sistemas baseados em conhecimento e a aprendizagem automática. O trabalho apresentado tem como objetivo a construção de um sistema para extrair informações a partir de textos contendo citações científicas (ou referências bibliográficas) através de uma abordagem baseada em aprendizagem automática. Dentre as diversas técnicas existentes, escolhemos tratar o problema através de uma abordagem híbrida, que combina o uso de técnicas de classificação de textos com os Modelos de Markov Escondidos (HMM). Esta combinação mostrou resultados superiores aos obtidos usando exclusivamente as técnicas de classificação e sua idéia básica é gerar com o uso das técnicas de classificação de textos para EI uma saída inicial para o sistema e refiná-la depois por meio de um HMM. Experimentos realizados com um conjunto de teste contendo 3000 referências resultaram em uma precisão de 87,48%.

In this thesis the Hidden Markov Model (HMM) is used in the process of map matching to investigate the accuracy for road segment map matching. A few HMM algorithms using a Global Navigation Satellite System (GNSS) receiver, odometer and gyroscope sensors are presented. The HMM algorithms are evaluated on four accuracy metrics. Two of these metrics have been seen in previous literature and captures road map match accuracy. The other have not been seen before and captures road segment accuracy. In the evaluation process a dataset is created by simulation to achieve positional ground truth for each sensor measurement. The accuracy distribution for different parts of the map matched trajectory is also evaluated. The result shows that HMM algorithms presented in previous literature, falls short to capture the accuracy for road segment map matching. The results further shows that by using less noisy sensors, as odometer and gyroscope, the accuracy for road segment map matching can be increased.

Doctor en Ingeniería Eléctrica / In this thesis an uncertainty weighting scheme for deep neural network-hidden Markov model (DNN-HMM) based automatic speech recognition (ASR) is proposed to increase discriminability in the decoding process. To this end, the DNN pseudo-log-likelihoods are weighted according to the uncertainty variance assigned to the acoustic observation. The results presented here suggest that substantial reduction in word error rate (WER) is achieved with clean training. Moreover, modelling the uncertainty propagation through the DNN is not required and no approximations for non linear activation functions are made. The presented method can be applied to any network topology that delivers log likelihood-like scores. It can be combined with any noise removal technique and adds a minimal computational cost. This technique was exhaustively evaluated and combined with uncertainty-propagation-based schemes for computing the pseudo-log-likelihoods and uncertainty variance at the DNN output. Two proposed methods optimized the parameters of the weighting function by leveraging the grid search either on a development database representing the given task or on each utterance based on discrimination metrics. Experiments with Aurora-4 task showed that, with clean training, the proposed weighting scheme can reduce WER by a maximum of 21% compared with a baseline system with spectral subtraction and uncertainty propagation using the unscented transform. Additionally, it is proposed to replace the classical black box integration of automatic speech recognition technology in human-robot interaction (HRI) applications with the incorporation of the HRI environment representation and modeling, and the robot and user states and contexts. Accordingly, this thesis focuses on the environment representation and modeling by training a DNN-HMM based automatic speech recognition engine combining clean utterances with the acoustic channel responses and noise that were obtained from an HRI testbed built with a PR2 mobile manipulation robot. This method avoids recording a training database in all the possible acoustic environments given an HRI scenario. In the generated testbed, the resulting ASR engine provided a WER that is at least 26% and 38% lower than publicly available speech recognition application programming interfaces (APIs) with the loudspeaker and human speakers testing databases, respectively, with a limited amount of training data. This thesis demonstrates that even state-of-the-art DNN-HMM based speech recognizers can benefit by combining systems for which the acoustic models have been trained using different feature sets. In this context, the complementarity of DNN-HMM based ASR systems trained with the same data set but with different signal representations is discussed. DNN fusion methods based on flat-weight combination, the minimization of mutual information and the maximization of discrimination metrics were proposed and tested. Schemes that consider the combination of ASR systems with lattice combination and minimum Bayes risk decoding were also evaluated and combined with DNN fusion techniques. The experimental results were obtained using a publicly-available naturally-recorded highly reverberant speech data. Significant improvements in WER were observed by combining DNN-HMM based ASR systems with different feature sets, obtaining relative improvements of 10% with two classifiers and 18% with four classifiers, without any tuning or a priori information of the ASR accuracy.

The master thesis is divided into two parts theoretical and practical. The theoretical part is focused on methods of analysis and detection of speech signals. In the practical part the system for isolated word recognition was created in Matlab. The system is speaker independent separately for men and women. Also two speech databases were created for further use in the aircraft cockpit. Tests and evaluations were performed even with added noise.

This thesis explores the use of head gesture recognition as an intuitive interface for computer interaction. This research presents a novel vision-based head gesture recognition system which utilizes contextual clues to reduce false positives. The system is used as a computer interface for answering dialog boxes. This work seeks to validate similar research, but focuses on using more efficient techniques using everyday hardware. A survey of image processing techniques for recognizing and tracking facial features is presented along with a comparison of several methods for tracking and identifying gestures over time. The design explains an efficient reusable head gesture recognition system using efficient lightweight algorithms to minimize resource utilization. The research conducted consists of a comparison between the base gesture recognition system and an optimized system that uses contextual clues to reduce false positives. The results confirm that simple contextual clues can lead to a significant reduction of false positives. The head gesture recognition system achieves an overall accuracy of 96% using contextual clues and significantly reduces false positives. In addition, the results from a usability study are presented showing that head gesture recognition is considered an intuitive interface and desirable above conventional input for answering dialog boxes. By providing the detailed design and architecture of a head gesture recognition system using efficient techniques and simple hardware, this thesis demonstrates the feasibility of implementing head gesture recognition as an intuitive form of interaction using preexisting infrastructure, and also provides evidence that such a system is desirable.

The meager offering of online commercial Swedish Automatic Speech Recognition ser-vices prompts the effort to develop a speech recognizer for Swedish using the open sourcetoolkit Kaldi and publicly available NST speech corpus. Using a previous Kaldi recipeseveral GMM-HMM models are trained and evaluated against commercial options toallow for reasoning of the performance of a customized solution for Automatic SpeechRecognition to that of commercial services. The evaluation takes both accuracy andcomputational speed into consideration. Initial results of the evaluation indicate a sys-tematic bias in the selected test set confirmed by a follow up investigative evaluation.The conclusion is that building a speech recognizer for Swedish using the NST corpusand Kaldi without expert knowledge is feasible but requires further work. / En taligenkännare för svenska utvecklas med målet att utvärdera hur en taligenkännareutvecklad med fritt tillgängliga verktyg står sig mot kommersiella taligenkänningstjänster.Verktyget som används är det öppna källkodsverktyget Kaldi och som träningsdataanvänds det offentligt tillgängliga talkorpuset för svenska från NST. De framtagna mod-ellerna jämförs mot kommersielt tillgängliga tjänster för taligenkänning på svenska.Tidiga resultat i jämförelsen indikerar ett systemiskt jäv i den valda testdata, vilketbekräftas av en uppföljande undersökande utvärdering. Slutsatsen av arbetet är attutsikterna att ta fram en taligenkännare för svenska är goda men kräver omfattandearbete.

This thesis describes the implementation of a speaker dependent connected-digit recognizer using continuous Hidden Markov Modeling (HMM). The speech recognition system was implemented using MATLAB and on the ADSP-2181, a digital signal processor manufactured by Analog Devices. Linear predictive coding (LPC) analysis was first performed on a speech signal to model the characteristics of the vocal tract filter. A 7 state continuous HMM with 4 mixture density components was used to model each digit. The Viterbi reestimation method was primarily used in the training phase to obtain the parameters of the HMM. Viterbi decoding was used for the recognition phase. The system was first implemented as an isolated word recognizer. Recognition rates exceeding 99% were obtained on both the MATLAB and the ADSP-2181 implementations. For continuous word recognition, several algorithms were implemented and compared. Using MATLAB, recognition rates exceeding 90% were obtained. In addition, the algorithms were implemented on the ADSP-2181 yielding recognition rates comparable to the MATLAB implementation. / Master of Science

La reconnaissance de mouvements naturels est de toute première importance dans la mise en oeuvre d’Interfaces Homme-Machine intelligentes et efficaces, utilisables de manière intuitive en environnement virtuel. En effet, elle permet à l’utilisateur d’agir de manière naturelle et au système de reconnaitre les mouvements corporel effectués tels qu’ils seraient perçu par un humain. Cette tâche est complexe, car elle demande de relever plusieurs défis : prendre en compte les spécificités du dispositif d’acquisition des données de mouvement, gérer la variabilité cinématique dans l’exécution du mouvement, et enfin gérer les différences morphologiques inter-individuelles, de sorte que les mouvements de tout nouvel utilisateur puissent être reconnus. De plus, de part la nature interactive des environnements virtuels, cette reconnaissancedoit pouvoir se faire en temps-réel, sans devoir attendre la fin du mouvement. La littérature scientifique propose de nombreuses méthodes pour répondre aux deux premiers défis mais la gestion de la variabilité morphologique est peu abordée. Dans cette thèse, nous proposons une description du mouvement permettant de répondre à cette problématique et évaluons sa capacité à reconnaitre les mouvements naturels d’un utilisateur inconnu. Enfin, nous proposons unenouvelle méthode permettant de tirer partie de cette représentation dans une reconnaissance précoce du mouvement / Recognition of natural movements is of utmost importance in the implementation of intelligent and effective Human-Machine Interfaces for virtual environments. It allows the user to behave naturally and the system to recognize its body movements in the same way a human might perceive it. This task is complex, because it addresses several challenges : take account of the specificities of the motion capture system, manage kinematic variability in motion performance, and finally take account of the morphological differences between individuals, so that actions of any new user can be recognized. Moreover, due to the interactive nature of virtual environments, this recognition must be achieved in real-time without waiting for the motion end. The literature offers many methods to meet the first two challenges. But the management of the morphological variability is not dealt. In this thesis, we propose a description of the movement to address this issue and we evaluate its ability to recognize the movements of an unknown user. Finally, we propose a new method to take advantage of this representation in early motion recognition