O aprendizado semi-supervisionado é um paradigma do aprendizado de máquina no qual a hipótese é induzida aproveitando tanto os dados rotulados quantos os dados não rotulados. Este paradigma é particularmente útil quando a quantidade de exemplos rotulados é muito pequena e a rotulação manual dos exemplos é uma tarefa muito custosa. Nesse contexto, foi proposto o algoritmo Cotraining, que é um algoritmo muito utilizado no cenário semi-supervisionado, especialmente quando existe mais de uma visão dos dados. Esta característica do algoritmo Cotraining faz com que a sua aplicabilidade seja restrita a domínios multi-visão, o que diminui muito o potencial do algoritmo para resolver problemas reais. Nesta dissertação, é proposto o algoritmo Co2KNN, que é uma versão mono-visão do algoritmo Cotraining na qual, ao invés de combinar duas visões dos dados, combina duas estratégias diferentes de induzir classificadores utilizando a mesma visão dos dados. Tais estratégias são chamados de k-vizinhos mais próximos (KNN) Local e Global. No KNN Global, a vizinhança utilizada para predizer o rótulo de um exemplo não rotulado é conformada por aqueles exemplos que contém o novo exemplo entre os seus k vizinhos mais próximos. Entretanto, o KNN Local considera a estratégia tradicional do KNN para recuperar a vizinhança de um novo exemplo. A teoria do Aprendizado Semi-supervisionado Baseado em Desacordo foi utilizada para definir a base teórica do algoritmo Co2KNN, pois argumenta que para o sucesso do algoritmo Cotraining, é suficiente que os classificadores mantenham um grau de desacordo que permita o processo de aprendizado conjunto. Para avaliar o desempenho do Co2KNN, foram executados diversos experimentos que sugerem que o algoritmo Co2KNN tem melhor performance que diferentes algoritmos do estado da arte, especificamente, em domínios mono-visão. Adicionalmente, foi proposto um algoritmo otimizado para diminuir a complexidade computacional do KNN Global, permitindo o uso do Co2KNN em problemas reais de classificação / Semi-supervised learning is a machine learning paradigm in which the induced hypothesis is improved by taking advantage of unlabeled data. Semi-supervised learning is particularly useful when labeled data is scarce and difficult to obtain. In this context, the Cotraining algorithm was proposed. Cotraining is a widely used semisupervised approach that assumes the availability of two independent views of the data. In most real world scenarios, the multi-view assumption is highly restrictive, impairing its usability for classifification purposes. In this work, we propose the Co2KNN algorithm, which is a one-view Cotraining approach that combines two different k-Nearest Neighbors (KNN) strategies referred to as global and local k-Nearest Neighbors. In the global KNN, the nearest neighbors used to classify a new instance are given by the set of training examples which contains this instance within its k-nearest neighbors. In the local KNN, on the other hand, the neighborhood considered to classify a new instance is the set of training examples computed by the traditional KNN approach. The Co2KNN algorithm is based on the theoretical background given by the Semi-supervised Learning by Disagreement, which claims that the success of the combination of two classifiers in the Cotraining framework is due to the disagreement between the classifiers. We carried out experiments showing that Co2KNN improves significatively the classification accuracy specially when just one view of training data is available. Moreover, we present an optimized algorithm to cope with time complexity of computing the global KNN, allowing Co2KNN to tackle real classification problems

Treinamento semi-supervisionado é uma metodologia de aprendizado de máquina que conjuga características de treinamento supervisionado e não-supervisionado. Ela se baseia no uso de bases semi-rotuladas (bases contendo dados rotulados e não-rotulados) para o treinamento de classificadores. A adição de dados não-rotulados, mais baratos e geralmente disponíveis em maior quantidade do que os dados rotulados, pode aumentar o desempenho e/ou baratear o custo de treinamento desses classificadores (a partir da diminuição da quantidade de dados rotulados necessários). Esta dissertação analisa duas estratégias para se executar treinamento semi-supervisionado, especificamente em Support Vector Machines (SVMs): formas direta e indireta. A estratégia direta é atualmente mais conhecida e estudada, e permite o uso de dados rotulados e não-rotulados, ao mesmo tempo, em tarefas de aprendizagem de classificadores. Entretanto, a inclusão de muitos dados não-rotulados pode tornar o treinamento demasiadamente lento. Já a estratégia indireta é mais recente, sendo capaz de agregar os benefícios do treinamento semi-supervisionado direto com tempos menores para o aprendizado de classificadores. Esta opção utiliza os dados não-rotulados para pré-processar a base de dados previamente à tarefa de aprendizagem do classificador, permitindo, por exemplo, a filtragem de eventuais ruídos e a reescrita da base em espaços de variáveis mais convenientes. Dentro do escopo da forma indireta, está a principal contribuição dessa dissertação: idealização, implementação e análise do algoritmo split learning. Foram obtidos ótimos resultados com esse algoritmo, que se mostrou eficiente em treinar SVMs de melhor desempenho e em períodos menores a partir de bases semi-rotuladas. / Semi-supervised learning is a machine learning methodology that mixes features of supervised and unsupervised learning. It allows the use of partially labeled databases (databases with labeled and unlabeled data) to train classifiers. The addition of unlabeled data, which are cheaper and generally more available than labeled data, can enhance the performance and/or decrease the costs of learning such classifiers (by diminishing the quantity of required labeled data). This work analyzes two strategies to perform semi-supervised learning, specifically with Support Vector Machines (SVMs): direct and indirect concepts. The direct strategy is currently more popular and studied; it allows the use of labeled and unlabeled data, concomitantly, in learning classifiers tasks. However, the addition of many unlabeled data can lead to very long training times. The indirect strategy is more recent; it is able to attain the advantages of the direct semi-supervised learning with shorter training times. This alternative uses the unlabeled data to pre-process the database prior to the learning task; it allows denoising and rewriting the data in better feature espaces. The main contribution of this Master thesis lies within the indirect strategy: conceptualization, experimentation, and analysis of the split learning algorithm, that can be used to perform indirect semi-supervised learning using SVMs. We have obtained promising empirical results with this algorithm, which is efficient to train better performance SVMs in shorter times from partially labeled databases.

A anotac~ao de papeis sem^anticos (APS) e uma tarefa do processamento de lngua natural (PLN) que permite analisar parte do signicado das sentencas atraves da detecc~ao dos participantes dos eventos (e dos eventos em si) que est~ao sendo descritos nelas, o que e essencial para que os computadores possam usar efetivamente a informac~ao codicada no texto. A maior parte das pesquisas desenvolvidas em APS tem sido feita para textos em ingl^es, considerando as particularidades gramaticais e sem^anticas dessa lngua, o que impede que essas ferramentas e resultados sejam diretamente transportaveis para outras lnguas como o portugu^es. A maioria dos sistemas de APS atuais emprega metodos de aprendizado de maquina supervisionado e, portanto, precisa de um corpus grande de senten cas anotadas com papeis sem^anticos para aprender corretamente a tarefa. No caso do portugu^es do Brasil, um recurso lexical que prov^e este tipo de informac~ao foi recentemente disponibilizado: o PropBank.Br. Contudo, em comparac~ao com os corpora para outras lnguas como o ingl^es, o corpus fornecido por este projeto e pequeno e, portanto, n~ao permitiria que um classicador treinado supervisionadamente realizasse a tarefa de anotac~ao com alto desempenho. Para tratar esta diculdade, neste trabalho emprega-se uma abordagem semissupervisionada capaz de extrair informac~ao relevante tanto dos dados anotados disponveis como de dados n~ao anotados, tornando-a menos dependente do corpus de treinamento. Implementa-se o algoritmo self-training com modelos de regress~ ao logstica (ou maxima entropia) como classicador base, para anotar o corpus Bosque (a sec~ao correspondente ao CETENFolha) da Floresta Sinta(c)tica com as etiquetas do PropBank.Br. Ao algoritmo original se incorpora balanceamento e medidas de similaridade entre os argumentos de um verbo especco para melhorar o desempenho na tarefa de classicac~ao de argumentos. Usando um benchmark de avaliac~ao implementado neste trabalho, a abordagem semissupervisonada proposta obteve um desempenho estatisticamente comparavel ao de um classicador treinado supervisionadamente com uma maior quantidade de dados anotados (80,5 vs. 82,3 de \'F IND. 1\', p > 0, 01) / Semantic role labeling (SRL) is a natural language processing (NLP) task able to analyze part of the meaning of sentences through the detection of the events they describe and the participants involved, which is essential for computers to eectively understand the information coded in text. Most of the research carried out in SRL has been done for texts in English, considering the grammatical and semantic particularities of that language, which prevents those tools and results to be directly transported to other languages such as Portuguese. Most current SRL systems use supervised machine learning methods and require a big corpus of sentences annotated with semantic roles in order to learn how to perform the task properly. For Brazilian Portuguese, a lexical resource that provides this type of information has recently become available: PropBank.Br. However, in comparison with corpora for other languages such as English, the corpus provided by that project is small and it wouldn\'t allow a supervised classier to perform the labeling task with good performance. To deal with this problem, in this dissertation we use a semi-supervised approach capable of extracting relevant information both from annotated and non-annotated data available, making it less dependent on the training corpus. We implemented the self-training algorithm with logistic regression (or maximum entropy) models as base classier to label the corpus Bosque (section CETENFolha) from the Floresta Sintá(c)tica with the PropBank.Br semantic role tags. To the original algorithm, we incorporated balancing and similarity measures between verb-specic arguments so as to improve the performance of the system in the argument classication task. Using an evaluation benchmark implemented in this research project, the proposed semi-supervised approach has a statistical comparable performance as the one of a supervised classier trained with more annotated data (80,5 vs. 82,3 de \'F IND. 1\', p > 0, 01).

A tarefa de classificação em Aprendizado de Máquina consiste da criação de modelos computacionais capazes de identificar automaticamente a classe de objetos pertencentes a um domínio pré-definido a partir de um conjunto de exemplos cuja classe é conhecida. Existem alguns cenários de classificação nos quais cada objeto pode estar associado não somente a uma classe, mas a várias classes ao mesmo tempo. Adicionalmente, nesses cenários denominados multirrótulo, as classes podem ser organizadas em uma taxonomia que representa as relações de generalização e especialização entre as diferentes classes, definindo uma hierarquia de classes, o que torna a tarefa de classificação ainda mais específica, denominada classificação hierárquica. Os métodos utilizados para a construção desses modelos de classificação são complexos e dependem fortemente da disponibilidade de uma quantidade expressiva de exemplos previamente classificados. Entretanto, para muitas aplicações é difícil encontrar um número significativo desses exemplos. Além disso, com poucos exemplos, os algoritmos de aprendizado supervisionado não são capazes de construir modelos de classificação eficazes. Nesses casos, é possível utilizar métodos de aprendizado semissupervisionado, cujo objetivo é aprender as classes do domínio utilizando poucos exemplos conhecidos conjuntamente com um número considerável de exemplos sem a classe especificada. Neste trabalho são propostos, entre outros, métodos que fazem uso do aprendizado semissupervisionado baseado em desacordo coperspectiva, tanto para a tarefa de classificação multirrótulo plana quanto para a tarefa de classificação hierárquica. São propostos, também, outros métodos que utilizam o aprendizado ativo com intuito de melhorar a performance de algoritmos de classificação semissupervisionada. Além disso, são propostos dois métodos para avaliação de algoritmos multirrótulo e hierárquico, os quais definem estratégias para identificação dos multirrótulos majoritários, que são utilizados para calcular os valores baseline das medidas de avaliação. Foi desenvolvido um framework para realizar a avaliação experimental da classificação hierárquica, no qual foram implementados os métodos propostos e um módulo completo para realizar a avaliação experimental de algoritmos hierárquicos. Os métodos propostos foram avaliados e comparados empiricamente, considerando conjuntos de dados de diversos domínios. A partir da análise dos resultados observa-se que os métodos baseados em desacordo não são eficazes para tarefas de classificação complexas como multirrótulo e hierárquica. Também é observado que o problema central de degradação do modelo dos algoritmos semissupervisionados agrava-se nos casos de classificação multirrótulo e hierárquica, pois, nesses casos, há um incremento nos fatores responsáveis pela degradação nos modelos construídos utilizando aprendizado semissupervisionado baseado em desacordo coperspectiva / In machine learning, the task of classification consists on creating computational models that are able to automatically identify the class of objects belonging to a predefined domain from a set of examples whose class is known a priori. There are some classification scenarios in which each object can be associated to more than one class at the same time. Moreover, in such multilabeled scenarios, classes can be organized in a taxonomy that represents the generalization and specialization relationships among the different classes, which defines a class hierarchy, making the classification task, known as hierarchical classification, even more specific. The methods used to build such classification models are complex and highly dependent on the availability of an expressive quantity of previously classified examples. However, for a large number of applications, it is difficult to find a significant number of such examples. Moreover, when few examples are available, supervised learning algorithms are not able to build efficient classification models. In such situations it is possible to use semi-supervised learning, whose aim is to learn the classes of the domain using a few classified examples in conjunction to a considerable number of examples with no specified class. In this work, we propose methods that use the co-perspective disagreement based learning approach for both, the flat multilabel classification and the hierarchical classification tasks, among others. We also propose other methods that use active learning, aiming at improving the performance of semi-supervised learning algorithms. Additionally, two methods for the evaluation of multilabel and hierarchical learning algorithms are proposed. These methods define strategies for the identification of the majority multilabels, which are used to estimate the baseline evaluation measures. A framework for the experimental evaluation of the hierarchical classification was developed. This framework includes the implementations of the proposed methods as well as a complete module for the experimental evaluation of the hierarchical algorithms. The proposed methods were empirically evaluated considering datasets from various domains. From the analysis of the results, it can be observed that the methods based on co-perspective disagreement are not effective for complex classification tasks, such as the multilabel and hierarchical classification. It can also be observed that the main degradation problem of the models of the semi-supervised algorithms worsens for the multilabel and hierarchical classification due to the fact that, for these cases, there is an increase in the causes of the degradation of the models built using semi-supervised learning based on co-perspective disagreement

Expert labeling, tagging, and assessment are far more costly than the processes of collecting raw data. Generative modeling is a very powerful tool to tackle this real-world problem. It is shown here how these models can be used to allow for semi-supervised learning that performs very well in label-deficient conditions. The foundation for the work in this dissertation is built upon visualizing generative models' latent spaces to gain deeper understanding of data, analyze faults, and propose solutions. A number of novel ideas and approaches are presented to improve single-label classification. This dissertation's main focus is on extending semi-supervised Deep Generative Models for solving the multi-label problem by proposing unique mathematical and programming concepts and organization. In all naive mixtures, using multiple labels is detrimental and causes each label's predictions to be worse than models that utilize only a single label. Examining latent spaces reveals that in many cases, large regions in the models generate meaningless results. Enforcing a priori independence is essential, and only when applied can multi-label models outperform the best single-label models. Finally, a novel learning technique called open-book learning is described that is capable of surpassing the state-of-the-art classification performance of generative models for multi-labeled, semi-supervised data sets.

Activity recognition falls in general area of pattern recognition, but it resides mainly in temporal domain which leads to distinctive characteristics. We provide an extensive survey over existing tools including FSM, HMM, BNT, DBN, SCFG and Symbolic Network Approach (PNF-network). These tools are inefficient to meet many of the requirements of activity recognition, leading to this work to develop a new graphical model: Propagation Net (P-Net). Many activities can be represented by a partially ordered set of temporal intervals, each of which corresponds to a primitive motion. Each interval has both temporal and logical constraints that control the duration of the interval and its relationship with other intervals. P-Net takes advantage of such fundamental constraints that it provides an graphical conceptual model to describe the human knowledge and an efficient computational model to facilitate recognition and learning. P-Nets define an exponentially large joint distribution that standard bayesian inference cannot handle. We devise two approximation algorithms to interpret a multi-dimensional observation sequence of evidence as a multi-stream propagation process through P-Net. First, Local Maximal Search Algorithm (LMSA) is constructed with polynomial complexity; Second, we introduce a particle filter based framework, Discrete Condensation (D-Condensation) algorithm, which samples the discrete state space more efficiently then original Condensation. To construct a P-Net based system, we need two parts: P-Net and the corresponding detector set. Given topology information and detector library, P-Net parameters can be extracted easily from a relatively small number of positive examples. To avoid the tedious process of manually constructing the detector library, we introduce semi-supervised learning framework to build P-Net and the corresponding detectors together. Furthermore, we introduce the Contrast Boosting algorithm that forces the detectors to be as different as possible but not necessary to be non-overlapping. The classification and learning ability of P-Nets are verified on three data sets: 1)vision tracked indoor activity data set; 2)vision tracked glucose monitor calibration data set; 3)sensor data set on simple weight-lifting exercise. Comparison with standard SCFG and HMM prove a P-Net based system is easier to construct and has a superior ability to classify complex human activity and detect anomaly.

m-Estimation represents a broad class of estimators, including least-squares and maximum likelihood, and is a widely used tool for statistical inference. Its successful application however, often requires negotiating physical resources for desired levels of accuracy. These limiting factors, which we abstractly refer as costs, may be computational, such as time-limited cluster access for parameter learning, or they may be financial, such as purchasing human-labeled training data under a fixed budget. This thesis explores these accuracy- cost tradeoffs by proposing a family of estimators that maximizes a stochastic variation of the traditional m-estimator. Such "stochastic m-estimators" (SMEs) are constructed by stitching together different m-estimators, at random. Each such instantiation resolves the accuracy-cost tradeoff differently, and taken together they span a continuous spectrum of accuracy-cost tradeoff resolutions. We prove the consistency of the estimators and provide formulas for their asymptotic variance and statistical robustness. We also assess their cost for two concerns typical to machine learning: computational complexity and labeling expense. For the sake of concreteness, we discuss experimental results in the context of a variety of discriminative and generative Markov random fields, including Boltzmann machines, conditional random fields, model mixtures, etc. The theoretical and experimental studies demonstrate the effectiveness of the estimators when computational resources are insufficient or when obtaining additional labeled samples is necessary. We also demonstrate that in some cases the stochastic m-estimator is associated with robustness thereby increasing its statistical accuracy and representing a win-win.

This dissertation presents the development of a semi-supervised incremental learning framework with a multi-view perspective for image concept modeling. For reliable image concept characterization, having a large number of labeled images is crucial. However, the size of the training set is often limited due to the cost required for generating concept labels associated with objects in a large quantity of images. To address this issue, in this research, we propose to incrementally incorporate unlabeled samples into a learning process to enhance concept models originally learned with a small number of labeled samples. To tackle the sub-optimality problem of conventional techniques, the proposed incremental learning framework selects unlabeled samples based on an expected error reduction function that measures contributions of the unlabeled samples based on their ability to increase the modeling accuracy. To improve the convergence property of the proposed incremental learning framework, we further propose a multi-view learning approach that makes use of multiple features such as color, texture, etc., of images when including unlabeled samples. For robustness to mismatches between training and testing conditions, a discriminative learning algorithm, namely a kernelized maximal- figure-of-merit (kMFoM) learning approach is also developed. Combining individual techniques, we conduct a set of experiments on various image concept modeling problems, such as handwritten digit recognition, object recognition, and image spam detection to highlight the effectiveness of the proposed framework.

In many real-world applications of supervised learning, only a limited number of labeled examples are available because the cost of obtaining high-quality examples is high. Even with a relatively large number of labeled examples, the learning problem may still suffer from limited supervision as the complexity of the prediction function increases. Therefore, learning with limited supervision presents a major challenge to machine learning. With the goal of supervision reduction, this thesis studies the representation, discovery and incorporation of extra input and output information in learning. Information about the input space can be encoded by regularization. We first design a semi-supervised learning method for text classification that encodes the correlation of words inferred from seemingly irrelevant unlabeled text. We then propose a multi-task learning framework with a matrix-normal penalty, which compactly encodes the covariance structure of the joint input space of multiple tasks. To capture structure information that is more general than covariance and correlation, we study a class of regularization penalties on model compressibility. Then we design the projection penalty, which encodes the structure information from a dimension reduction while controlling the risk of information loss. Information about the output space can be exploited by error correcting output codes. Using the composite likelihood view, we propose an improved pairwise coding for multi-label classification, which encodes pairwise label density (as opposed to label comparisons) and decodes using variational methods. We then investigate problemdependent codes, where the encoding is learned from data instead of being predefined. We first propose a multi-label output code using canonical correlation analysis, where predictability of the code is optimized. We then argue that both discriminability and predictability are critical for output coding, and propose a max-margin formulation that promotes both discriminative and predictable codes. We empirically study our methods in a wide spectrum of applications, including document categorization, landmine detection, face recognition, brain signal classification, handwritten digit recognition, house price forecasting, music emotion prediction, medical decision, email analysis, gene function classification, outdoor scene recognition, and so forth. In all these applications, our proposed methods for encoding input and output information lead to significantly improved prediction performance.

Many automotive safety applications in modern cars make use of cameras and object detection to analyze the surrounding environment. Pedestrians, animals and other vehicles can be detected and safety actions can be taken before dangerous situations arise. To detect occurrences of the different objects, these systems are traditionally trained to learn a classification model using a set of images that carry labels corresponding to their content. To obtain high performance with a variety of object appearances, the required amount of data is very large. Acquiring unlabeled images is easy, while the manual work of labeling is both time-consuming and costly. Semi-supervised learning refers to methods that utilize both labeled and unlabeled data, a situation that is highly desirable if it can lead to improved accuracy and at the same time alleviate the demand of labeled data. This has been an active area of research in the last few decades, but few studies have investigated the performance of these algorithms in larger systems. In this thesis, we investigate if and how semi-supervised learning can be used in a large-scale pedestrian detection system. With the area of application being automotive safety, where real-time performance is of high importance, the work is focused around boosting classifiers. Results are presented on a few publicly available UCI data sets and on a large data set for pedestrian detection captured in real-life traffic situations. By evaluating the algorithms on the pedestrian data set, we add the complexity of data set size, a large variety of object appearances and high input dimension. It is possible to find situations in low dimensions where an additional set of unlabeled data can be used successfully to improve a classification model, but the results show that it is hard to efficiently utilize semi-supervised learning in large-scale object detection systems. The results are hard to scale to large data sets of higher dimensions as pair-wise computations are of high complexity and proper similarity measures are hard to find.