Libros y Capítulos de libros
URI permanente para esta colección
Examinar
Examinando Libros y Capítulos de libros por Materia "14 - I+D en relación con la Ingeniería"
Mostrando 1 - 3 de 3
Resultados por página
Opciones de ordenación
Ítem A new dynamic, secondary-memory metric index(Red de Universidades Nacionales con carreras en Informática (RedUNCI); Universidad Nacional de la Plata, 2024-10-07) Paredes, RodrigoMetric space searching addresses the problem of efficient similarity searching in many applications. Although promising, the metric space approach is still immature in several aspects that are well established in traditional databases. Particularly, most indexing schemes are not dynamic, that is, few of them tolerate insertion of elements at reasonable cost over an existing index with none or mild performance degrading; and even less of them work efficiently in secondary memory. The List of Clusters (LC) is a competitive index in main memory. We introduce a new dynamic, secondary-memory variant of the LC. Our new index handles well the secondary memory scenario and is competitive with the state of the art, becoming a useful alternative in a wide range of database applications. Also, our ideas are applicable to other secondary-memory indexes, where it is possible to control the disk page occupation.Ítem BOLDSC: A New Dynamic, Secondary-Memory Metric Index(Springer Nature, 2025-10-01) Paredes, RodrigoMetric space searching addresses the problem of efficient similarity searching across diverse applications, in particular for non-structured objects, for instance, natural language or images. Although promising, this approach is still immature in several aspects that are well-established in traditional databases. Particularly, most indexing schemes are not dynamic, as they cannot efficiently handle insertions over an ongoing index without significant performance degradation. Moreover, very few of them work efficiently in secondary memory. The List of Clusters (LC) has proven to be a competitive index in main memory due to its simplicity and good search performance in high dimensional metric spaces. We introduce a new dynamic, secondary-memory LC variant. Our new index efficiently handles the secondary memory scenario and achieves competitive search and insertion times compared to the state-of-the-art, making it a practical alternative for large-scale database applications. Also, our ideas are applicable to other secondary-memory indexes, where it is possible to control the disk page occupation.Ítem Negative sampling for triplet-based loss: improving representation in self-supervised representation learning(Springer, 2024-11-17) Goyo, Manuel AlejandroSignificant strides have been made in artificial neural networks across various fields, necessitating extensive labeled data for effective training. However, the acquisition of such annotated data is both costly and labor-intensive. To address this challenge, Self-Supervised Representation Learning (SSRL) has emerged as a promising solution. One prominent SSRL method, Contrastive Self-Supervised Learning (CSL), enhances feature representations by discerning similarities and differences among samples in the feature space. Yet, accurately identifying dissimilar samples remains a persistent issue, limiting CSL’s effectiveness. In response, an innovative enhancement to CSL is proposed in this paper. Explicit negative sampling strategies using a binary classification algorithm within the feature space are introduced to distinguish between similar and dissimilar features precisely. Additionally, Triplet Loss, originally designed for tasks such as person re-identification and face recognition, is incorporated to further refine feature learning. Experimental evaluations on the CIFAR-10 and SVHN datasets validate the proposed method’s superiority in content-based image retrieval (CBIR) and classification tasks. Significant improvements are demonstrated in metrics such as mean average precision (MAP), accuracy, recall, precision, and F1-score compared to existing techniques. This framework contributes to the advancement of SSRL by enabling scalable neural network training on large datasets with minimal annotation, effectively bridging the gap between supervised and unsupervised learning paradigms.