Doutorado em Ciência da Computação
URI Permanente para esta coleção
Navegar
Navegando Doutorado em Ciência da Computação por Por Orientador "Soares, Anderson da Silva"
Agora exibindo 1 - 3 de 3
Resultados por página
Opções de Ordenação
Item Future-Shot: Few-Shot Learning to tackle new labels on high-dimensional classification problems(Universidade Federal de Goiás, 2024-02-23) Camargo, Fernando Henrique Fernandes de; Soares, Anderson da Silva; http://lattes.cnpq.br/1096941114079527; Soares, Anderson da Silva; Galvão Filho, Arlindo Rodrigues; Vieira, Flávio Henrique Teles; Gomes, Herman Martins; Lotufo, Roberto de AlencarThis thesis introduces a novel approach to address high-dimensional multiclass classification challenges, particularly in dynamic environments where new classes emerge. Named Future-Shot, the method employs metric learning, specifically triplet learning, to train a model capable of generating embeddings for both data points and classes within a shared vector space. This facilitates efficient similarity comparisons using techniques like k-nearest neighbors (\acrshort{knn}), enabling seamless integration of new classes without extensive retraining. Tested on lab-of-origin prediction tasks using the Addgene dataset, Future-Shot achieves top-10 accuracy of $90.39\%$, surpassing existing methods. Notably, in few-shot learning scenarios, it achieves an average top-10 accuracy of $81.2\%$ with just $30\%$ of the data for new classes, demonstrating robustness and efficiency in adapting to evolving class structuresItem Classificação de cenas utilizando a análise da aleatoriedade por aproximação da complexidade de Kolmogorov(Universidade Federal de Goiás, 2020-03-15) Feitosa, Rafael Divino Ferreira; Soares, Anderson da Silva; http://lattes.cnpq.br/1096941114079527; Soares, Anderson da Silva; Delbem, Alexandre Cláudio Botazzo; Soares, Fabrízzio Alphonsus Alves de Melo Nunes; Laureano, Gustavo Teodoro; Costa, Ronaldo Martins daIn many pattern recognition problems, discriminant features are unknown and/or class boundaries are not well defined. Several studies have used data compression to discover knowledge, without features extraction and selection. The basic idea is two distinct objects can be grouped as similar, if the information content of one explains, in a significant way, the information content of the other. However, compressionbased techniques are not efficient for images, as they disregard the semantics present in the spatial correlation of two-dimensional data. A classifier is proposed for estimates the visual complexity of scenes, namely Pattern Recognition by Randomness (PRR). The operation of the method is based on data transformations, which expand the most discriminating features and suppress details. The main contribution of the work is the use of randomness as a measure discrimination. The approximation between scenes and trained models, based on representational distortion, promotes a lossy compression process. This loss is associated with irrelevant details, when the scene is reconstructed with the representation of true class, or with the information degradation, when it is reconstructed with divergent representations. The more information preserved, the greater the randomness of the reconstruction. From the mathematical point of view, the method is explained by two main measures in the U-dimensional plane: intersection and dispersion. The results yielded accuracy of 0.6967, for a 12-class problem, and 0.9286 for 7 classes. Compared with k-NN and a data mining toolkit, the proposed classifier was superior. The method is capable of generating efficient models from few training samples. It is invariant for vertical and horizontal reflections and resistant to some geometric transformations and image processing.Item Avaliação da qualidade da sintetização de fala gerada por modelos de redes neurais profundas(Universidade Federal de Goiás, 2023-05-26) Oliveira, Frederico Santos de; Soares, Anderson da Silva; http://lattes.cnpq.br/1096941114079527; Soares, Anderson da Silva; Aluisio, Sandra Maria; Duarte, Julio Cesar; Laureano, Gustavo Teodoro; Galvão Filho, Arlindo RodriguesWith the emergence of intelligent personal assistants, the need for high-quality conversational interfaces has increased. While text-based chatbots are popular, the development of voice interfaces is equally important. However, the primary method for evaluating voice-based conversational models is mainly done through Mean Opinion Score (MOS), which relies on a manual and subjective process. In this context, this thesis aims to contribute with a new methodology for evaluating voice-based conversational interfaces, with a case study specifically conducted in Brazilian Portuguese. The proposed methodology includes an architecture for predicting the quality of synthesized speech in Brazilian Portuguese, correlated with MOS. To evaluate the proposed methodology, this work included training Text-to-Speech models to create the dataset called BRSpeechMOS. Details about the creation of this dataset are presented, along with a qualitative and quantitative analysis of it. A series of experiments were conducted to train various architectures using the BRSpeechMOS dataset. The architectures used are based on supervised and self-supervised learning. The results obtained confirm the hypothesis raised that pre-trained models on voice processing tasks such as speaker verification and automatic speech recognition produce suitable acoustic representations for the task of predicting speech quality, contributing to the advancement of the state of the art in the development of evaluation methodologies for conversational models.