O cérebro humano é formado por um conjunto de neurônios de diferentes tipos, cada um com sua especialidade. A combinação destes diferentes tipos de neurônios é um dos aspectos responsáveis pelo desempenho apresentado pelo cérebro na realização de várias tarefas. Redes Neurais Artificiais são técnicas computacionais que apresentam um modelo matemático inspirado no sistema nervoso e que adquirem conhecimento através da experiência. Uma alternativa para melhorar o desempenho das Redes Neurais Artificiais é a utilização de técnicas de Combinação de Classificadores. Estas técnicas de combinação exploram as diferenças e as semelhanças das redes para a obtenção de resultados melhores. Dentre as principais aplicações de Redes Neurais Artificiais está o Reconhecimento de Padrões. Neste trabalho, foram utilizadas técnicas de Combinação de Classificadores para a combinação de Redes Neurais Artificiais em problemas de Reconhecimento de Padrões. / The human brain is formed by neurons of different types, each one with its own speciality. The combination of theses different types of neurons is one of the main features responsible for the brain performance in severa! tasks. Artificial Neural Networks are computation technics whose mathematical model is based on the nervous system and learns new knowledge by experience. An alternative to improve the performance of Artificial Neural Networks is the employment of Classifiers Combination techniques. These techniques of combination explore the difference and the similarity of the networks to achieve better performance. The main application of Artificial Neural Networks is Pattern Recognition. In this work, Classifiers Combination techniques were utilized to combine Artificial Neural Networks to solve Pattern Recognition problems.

Les tags constituent un outil très utile pour indexer des documents multimédias. Cette thèse de doctorat s’intéresse au tagging automatique, c’est à dire l’association automatique par un algorithme d’un ensemble de tags à chaque morceau. Nous utilisons des techniques de boosting pour réaliser un apprentissage prenant mieux en compte la richesse de l’information exprimée par la musique. Un algorithme de boosting est proposé, afin d’utiliser conjointement des descriptions de morceaux associées à des extraits de différentes durées. Nous utilisons cet algorithme pour fusionner de nouvelles descriptions, appartenant à différents niveaux d’abstraction. Enfin, un nouveau cadre d’apprentissage est proposé pour le tagging automatique, qui prend mieux en compte les subtilités des associations entre les tags et les morceaux. / Tags constitute a very useful tool for multimedia document indexing. This PhD thesis deals with automatic tagging, which consists in associating a set of tags to each song automatically, using an algorithm. We use boosting techniques to design a learning which better considers the complexity of the information expressed by music. A boosting algorithm is proposed, which can jointly use song descriptions associated to excerpts of different durations. This algorithm is used to fuse new descriptions, which belong to different abstraction levels. Finally, a new learning framework is proposed for automatic tagging, which better leverages the subtlety ofthe information expressed by music.

No description available.

Emotional intelligence has a prominent role in education, health care, and day to day interaction. With the increasing use of computer technology, computers are interacting with more and more individuals. This interaction provides an opportunity to increase knowledge about human emotion for human consumption, well-being, and improved computer adaptation. This thesis explores the efficacy of using up to four different sensors in three domains for computational affect detection. We first consider computer-based education, where a collection of four sensors is used to detect student emotions relevant to learning, such as frustration, confidence, excitement and interest while students use a computer geometry tutor. The best classier of each emotion in terms of accuracy ranges from 78% to 87.5%. We then use voice data collected in a clinical setting to differentiate both gender and culture of the speaker. We produce classifiers with accuracies between 84% and 94% for gender, and between 58% and 70% for American vs. Asian culture, and we find that classifiers for distinguishing between four cultures do not perform better than chance. Finally, we use video and audio in a health care education scenario to detect students' emotions during a clinical simulation evaluation. The video data provides classifiers with accuracies between 63% and 88% for the emotions of confident, anxious, frustrated, excited, and interested. We find the audio data to be too complex to single out the voice source of the student by automatic means. In total, this work is a step forward in the automatic computational detection of affect in realistic settings.

Radio-frequency (RF) fingerprinting is a process that uses the minute inconsistencies among manufactured radio transmitters to identify wireless devices. Coupled with location fingerprinting, which is a machine learning technique to locate devices based on their radio signals, it can uniquely identify and locate both trusted and rogue wireless devices transmitting over the air. This can have wide-ranging applications for the Internet of Things, security, and networking fields. To contribute to this effort, this research first builds a software-defined radio (SDR) testbed to collect an RF dataset over LTE and WiFi channels. The developed testbed consists of both hardware which are receivers with multiple antennas and software which performs signal preprocessing. Several features that can be used for RF device fingerprinting and location fingerprinting, including received signal strength indicator and channel state information, are also extracted from the signals. With the developed dataset, several data-driven machine learning algorithms have been implemented and tested for fingerprinting performance evaluation. Overall, experimental results show promising performance with a radio fingerprinting accuracy above 90\% and device localization within 1.10 meters.

No

The IEEE 802.11 based Wireless LAN (WLAN) has become increasingly ubiquitous in recent years. However, due to the broadcast nature of wireless communication, attackers can exploit the existing vulnerabilities in IEEE 802.11 to launch various types of attacks in wireless and wired networks. This thesis presents a statistical based hybrid Intrusion Detection System (IDS) for Rogue Access Point (RAP) detection, which employs distributed monitoring devices to monitor on 802.11 link layer activities and a centralized detection module at a gateway router to achieve higher accuracy in detection of rogue devices. This detection approach is scalable, non-intrusive and does not require any specialized hardware. It is designed to utilize the existing wireless LAN infrastructure and is independent of 802.11a/b/g/n. It works on passive monitoring of wired and wireless traffic, and hence is easy to manage and maintain. In addition, this approach requires monitoring a smaller number of packets for detection as compared to other detection approaches in a heterogeneous network comprised of wireless and wired subnets. Centralized detection is done at a gateway router by differentiating wired and wireless TCP traffic using Weighted Sequential Hypothesis Testing on inter-arrival time of TCP ACK-pairs. A decentralized module takes care of detection of MAC spoofing and totally relies on 802.11 beacon frames. Detection is done through analysis of the clock skew and the Received Signal Strength (RSS) as fingerprints using a naïve Bayes classifier to detect presence of rogue APs. Analysis of the system and extensive experiments in various scenarios on a real system have proven the efficiency and accuracy of the approach with few false positives/negatives and low computational and storage overhead. / Master of Science

Microbiome research has become a ubiquitous component of contemporary clinical research, with potential to uncover associations between microbiome composition and disease. With microbiome data becoming more prevalent, the need to understand how to analyse such data is increasingly important. One complicating property of microbiome data is that it is inherently compositional and thus constrained to simplex-space; because of this, it cannot be analysed directly using conventional statistical methods. In this paper, we transform the compositional data in order to lift the simplex-constraint, and then investigate the viability of applying conventional statistical methods to the data. Using a high-dimensional data set containing gut-microbiome samples from Parkinson's- and control patients, we first transform the raw data to centred log-ratio scale, and then use permutational multivariate analysis of variance (PERMANOVA) to test if there are differences between the two groups with respect to bacterial abundances. We then employ three machine learning classifiers -- Logistic regression, XGBoost, and Random Forest -- and evaluate their performance on the transformed data. The results from PERMANOVA indicate that gut-microbiome composition in the patients with Parkinson's disease indeed differ from that in the control individuals. The Random Forest method achieves the highest classification accuracy, followed by XGBoost, while logistic regression performs poorly, questioning its viability in analysis of high-dimensional compositional microbiome data. We find four bacterial species of high importance for the classification: Prevotella copri, Prevotella sp. CAG 520, Akkermansia muciniphila, and Butyricimonas virosa, where the first three have been previously mentioned in the Parkinson's literature.

What explains the shape of Zauzou noun phrases? Zauzou (Trans-Himalayan, China) noun phrases exhibit considerable diversity in both the choice of the phrase's primary reference type, and the presence of classifiers, demonstratives, modifiers, and case marking. This investigation uses a large, previously existing Zauzou textual corpus. The corpus was annotated for variables hypothesized to predict the variation in noun phrase form. Syntactic variables investigated include word order, subordination, subordinate role, and a new variable called "loneliness." Participant semantic variables include thematic role, agency, and affectedness. Referential semantic variables include boundedness, number, and animacy. The information packaging variable investigated is accessibility. Statistical analysis of the corpus revealed that case marking was predicted using a variable called "loneliness." This is where a multivalent verb has only one argument that is explicitly referenced in the clause. Lonely noun phrases are more likely to be case marked. The role of loneliness in motivating case marking confirms that disambiguation can be an explanation for differential case marking. Animacy and accessibility are important predictors of noun phrase weight. Overall, high animacy and high accessibility correspond to reduced noun phrase weight. Agency and thematic role were also significant variables. The Zauzou data makes clear that speech act participants occupy a unique role in the animacy hierarchy. Speech act participants are often unexpectedly light upon first mention, being referred to with a pronoun or zero anaphor. They are often unexpectedly heavy while highly activated, remaining a pronoun instead of reducing to a zero anaphor. Zauzou, like Mandarin and Cantonese, allows classifiers to be used with a noun but without a numeral. In Mandarin, this construction is used only with new or generic noun phrases. In Cantonese, this construction can be used with noun phrases of any accessibility value. Zauzou occupies a unique intermediate position. In Zauzou, a noun with bare noun phrase can occur with new or old noun phrases, but rarely with active ones. This thesis provides evidence for the importance of text corpora. Using a corpus allowed for the simultaneous inclusion of many variables as well as the consideration of genre effects. In addition, the annotated corpus produced in this investigation is an important output; it is available in the supplemental materials accompanying this thesis.

Many real-world problems like hand-written digit recognition or semantic scene classification are treated as multiclass or multi-label classification prob-lems. Solutions to these problems using support vector machines (SVMs) are well studied in literature. In this work, we focus on building sparse max-margin classifiers for multiclass and multi-label classification. Sparse representation of the resulting classifier is important both from efficient training and fast inference viewpoints. This is true especially when the training and test set sizes are large.Very few of the existing multiclass and multi-label classification algorithms have given importance to controlling the sparsity of the designed classifiers directly. Further, these algorithms were not found to be scalable. Motivated by this, we propose new formulations for sparse multiclass and multi-label classifier design and also give efficient algorithms to solve them. The formulation for sparse multi-label classification also incorporates the prior knowledge of label correlations. In both the cases, the classification model is designed using a common set of basis vectors across all the classes. These basis vectors are greedily added to an initially empty model, to approximate the target function. The sparsity of the classifier can be controlled by a user defined parameter, dmax which indicates the max-imum number of common basis vectors. The computational complexity of these algorithms for multiclass and multi-label classifier designisO(lk2d2 max), Where l is the number of training set examples and k is the number of classes. The inference time for the proposed multiclass and multi-label classifiers is O(kdmax). Numerical experiments on various real-world benchmark datasets demonstrate that the proposed algorithms result in sparse classifiers that require lesser number of basis vectors than required by state-of-the-art algorithms, to attain the same generalization performance. Very small value of dmax results in significant reduction in inference time. Thus, the proposed algorithms provide useful alternatives to the existing algorithms for sparse multiclass and multi-label classifier design.