Submitted by Marcele Costal de Castro (costalcastro@gmail.com) on 2017-09-20T18:26:27Z No. of bitstreams: 1 JOSE CARLOS SOARES DE ALMEIDA M.PDF: 406246 bytes, checksum: 8c9fa9ca253d578b9e0f7b938d783030 (MD5) / Made available in DSpace on 2017-09-20T18:26:27Z (GMT). No. of bitstreams: 1 JOSE CARLOS SOARES DE ALMEIDA M.PDF: 406246 bytes, checksum: 8c9fa9ca253d578b9e0f7b938d783030 (MD5) Previous issue date: 2001-02 / Este trabalho desenvolve um método de identificação de transientes com abordagem possibilística, utilizando algoritmo genético para a otimização do número de centroides das classes que representam os transientes. A ideia básica do método é otimizar a partição do espaço de busca, gerando um número mínimo de subclasses, definidas como subconjuntos de classes dentro de uma partição, cujos centroides consigam distinguir as classes com o máximo de acerto nas classificações. A interpretação das subclasses como conjuntos nebulosos e a abordagem possibilística forneceram um heurística para estabelecer zonas de influência dos centroides, possibilitando a intenção da resposta “Não Sei” para transientes desconhecidos, isto é, não pertencentes ao conjunto de treinamentos.

<p> Speaker age and gender classification is one of the most challenging problems in speech processing. Recently with developing technologies, identifying a speaker age and gender has become a necessity for speaker verification and identification systems such as identifying suspects in criminal cases, improving human-machine interaction, and adapting music for awaiting people queue. Although many studies have been carried out focusing on feature extraction and classifier design for improvement, classification accuracies are still not satisfactory. The key issue in identifying speaker&rsquo;s age and gender is to generate robust features and to design an in-depth classifier. Age and gender information is concealed in speaker&rsquo;s speech, which is liable for many factors such as, background noise, speech contents, and phonetic divergences.</p><p> In this work, different methods are proposed to enhance the speaker age and gender classification based on the deep neural networks (DNNs) as a feature extractor and classifier. First, a model for generating new features from a DNN is proposed. The proposed method uses the Hidden Markov Model toolkit (HTK) tool to find tied-state triphones for all utterances, which are used as labels for the output layer in the DNN. The DNN with a bottleneck layer is trained in an unsupervised manner for calculating the initial weights between layers, then it is trained and tuned in a supervised manner to generate transformed mel-frequency cepstral coefficients (T-MFCCs). Second, the shared class labels method is introduced among misclassified classes to regularize the weights in DNN. Third, DNN-based speakers models using the SDC feature set is proposed. The speakers-aware model can capture the characteristics of the speaker age and gender more effectively than a model that represents a group of speakers. In addition, AGender-Tune system is proposed to classify the speaker age and gender by jointly fine-tuning two DNN models; the first model is pre-trained to classify the speaker age, and second model is pre-trained to classify the speaker gender. Moreover, the new T-MFCCs feature set is used as the input of a fusion model of two systems. The first system is the DNN-based class model and the second system is the DNN-based speaker model. Utilizing the T-MFCCs as input and fusing the final score with the score of a DNN-based class model enhanced the classification accuracies. Finally, the DNN-based speaker models are embedded into an AGender-Tune system to exploit the advantages of each method for a better speaker age and gender classification.</p><p> The experimental results on a public challenging database showed the effectiveness of the proposed methods for enhancing the speaker age and gender classification and achieved the state of the art on this database.</p><p>

<p> Despite the extraordinary success of deep learning on diverse problems, these triumphs are too often confined to large, clean datasets and well-defined objectives. Face recognition systems train on millions of perfectly annotated images. Commercial speech recognition systems train on thousands of hours of painstakingly-annotated data. But for applications addressing human activity, data can be noisy, expensive to collect, and plagued by missing values. In electronic health records, for example, each attribute might be observed on a different time scale. Complicating matters further, deciding precisely what objective warrants optimization requires critical consideration of both algorithms and the application domain. Moreover, deploying human-interacting systems requires careful consideration of societal demands such as safety, interpretability, and fairness.</p><p> The aim of this thesis is to address the obstacles to mining temporal patterns in human activity data. The primary contributions are: (1) the first application of RNNs to multivariate clinical time series data, with several techniques for bridging long-term dependencies and modeling missing data; (2) a neural network algorithm for forecasting surgery duration while simultaneously modeling heteroscedasticity; (3) an approach to quantitative investing that uses RNNs to forecast company fundamentals; (4) an exploration strategy for deep reinforcement learners that significantly speeds up dialogue policy learning; (5) an algorithm to minimize the number of catastrophic mistakes made by a reinforcement learner; (6) critical works addressing model interpretability and fairness in algorithmic decision-making.</p><p>

<p> Deep learning networks in the literature traditionally only used a single input modality (or data stream). Integrating multiple modalities into deep learning networks with the goal of correlating extracted features was a major issue. Traditional methods involved treating each modality separately and then writing custom code to combine the extracted features.</p><p> Current solutions for small numbers of modalities (three or less) showed there are multiple architectures for modality integration. With an increase in the number of modalities, the &ldquo;curse of dimensionality&rdquo; affects the performance of the system. The research showed current methods for larger scale integrations required separate, custom created modules with another integration layer outside the deep learning network. These current solutions do not scale well nor provide good generalized performance. This research report studied architectures using multiple modalities and the creation of a scalable and efficient architecture.</p>

Abstract Not Available.

There are many tools available today to help software engineers search in source code systems. It is often the case, however, that there is a gap between what people really want to find and the actual query strings they specify. This is because a concept in a software system may be represented by many different terms, while the same term may have different meanings in different places. Therefore, software engineers often have to guess as they specify a search, and often have to repeatedly search before finding what they want. To alleviate the search problem, this thesis describes a study of what we call intelligent search techniques as implemented in a software exploration environment, whose purpose is to facilitate software maintenance. We propose to utilize some information retrieval techniques to automatically apply transformations to the query strings. The thesis first introduces the intelligent search techniques used in our study, including abbreviation concatenation and abbreviation expansion. Then it describes in detail the rating algorithms used to evaluate the query results' similarity to the original query strings. Next, we describe a series of experiments we conducted to assess the effectiveness of both the intelligent search methods and our rating algorithms. Finally, we describe how we use the analysis of the experimental results to recommend an effective combination of searching techniques for software maintenance, as well as to guide our future research.

Most of the search methods used in AI are inflexible. Interactive search is a new kind of search in which the search system can communicate and cooperate with external agents. There are two kinds of agents: human agents and non-human agents. Through interaction with human agents (man-machine interaction), the search system can make use of the human talent of judging the quality of a solution. Through interaction with non-human agents (machine-machine interaction), the search system can automatically exploit knowledge from its environment. An interactive search system has the ability to take advice from external agents. The ordinary non-interactive search models are the special instances of interactive search when the advice sequences are empty. We are investigating a particular kind of Interactive Search, IHGT (Interactive Hierarchical Generate and Test) search, which is established by introducing interactive ability into HGT (Hierarchical Generate and Test) search. To make HGT search interactive, we created an editor called GE (Generator Editor). GE was implemented in Prolog. GE is a bottom level language shell outside the HGT search model which translates advice into dynamic changes of all the three search factors. (Abstract shortened by UMI.)

This thesis demonstrates how the power of symbolic processing can be exploited in the learning of low level control functions. It proposes a novel hybrid architecture with a tight coupling between a variant of symbolic planning and reinforcement learning. This architecture combines the strengths of the function approximation of subsymbolic learning with the more abstract compositional nature of symbolic learning. The former is able to represent mappings of world states to actions in an accurate way. The latter allows a more rapid solution to problems by exploiting structure within the domain. A control function is learnt over time through interaction with the world. Symbols are attached to features in the functions. The symbolic attachments act as anchor points used to transform the function of a previously learnt task to that of a new task. The solution of more complex tasks is achieved through composing simpler functions, using the symbolic attachments to determine the composition. The result is used as the initial control function of the new task and then modified through further learning. This is shown to produce a significant speed up over basic reinforcement learning.

In many domains, characterizations of a given attribute are imprecise, uncertain and incomplete in the available learning examples. The definitions of classes may be vague. Learning systems are frequently forced to deal with such uncertainty. Traditional learning systems are designed to work in the domains where imprecision and uncertainty in the data are absent. Those learning systems are limited because of their impossibility to cope with uncertainty--a typical feature of real-world data. In this thesis, we developed a fuzzy learning system which combines inductive learning with a fuzzy approach to solve problems arising in learning tasks in the domains affected by uncertainty and vagueness. Based on Fuzzy Logic, rather than pure First Order Logic used in FOIL, this system extends FOIL with learning fuzzy logic relation from both imprecise examples and background knowledge represented by Fuzzy Prolog. The classification into the positive and negative examples is allowed to be a degree (of positiveness or negativeness) between 0 and 1. The values of a given attribute in examples need not to be the same type. Symbolic and continuous data can exist in the same attribute, allowing for fuzzy unification (inexact matching). An inductive learning problem is formulated as to find a fuzzy logic relation with a degree of truth, in which a fuzzy gain calculation method is used to guide heuristic search. The Fuzzy FOIL's ability of learning the required fuzzy logic relations and dealing with vague data enhances FOIL's usefulness.

This thesis addresses the problem of learning concept descriptions that are interpretable, or explainable. Explainability is understood as the ability to justify the learned concept in terms of the existing background knowledge. The starting point for the work was an existing system that would induce only fully explainable rules. The system performed well when the model used during induction was complete and correct. In practice, however, models are likely to be imperfect, i.e. incomplete and incorrect. We report here a new approach that achieves explainability with imperfect models. The basis of the system is the standard inductive search driven by an accuracy-oriented heuristic, biased towards rule explainability. The bias is abandoned when there is heuristic evidence that a significant loss of accuracy results from constraining the search to explainable rules only. The users can express their relative preference for accuracy vs. explainability. Experiments with the system indicate that, even with a partially incomplete and/or incorrect model, insisting on explainability results in only a small loss of accuracy. We also show how the new approach described can repair a faulty model using evidence derived from data during induction.