Advanced search
1 file | 7.44 MB

Supervised distance metric learning for pattern recognition

Bac Nguyen Cong (UGent)
(2019)
Author
Promoter
(UGent) and Carlos Morell Pérez
Organization
Abstract
Much like in other modeling disciplines does the distance metric used (a measure for dissimilarity) play an important role in the growing field of machine learning. Often, predefined distance metrics (e.g. the Euclidean one) are used to perform such measurement. Unfortunately, most of them ignore any statistical properties that might be estimated from the data. The notion of a good distance metric changes when one moves from one domain to another. For instance, in the problem of computing the dissimilarity for human images, two images could be considered as being similar due to one of the following reasons, the two images are taken from two persons with the same gender, the same age, or the same race. Clearly, it is difficult to use the same distance metric for gender, age, and race since two images might be similar in one case, while being dissimilar in the other case. For this reason, most research efforts have been devoted to automatically learn a good distance metric from data. Depending on the availability of training data, distance metric learning methods can be divided into three categories: supervised, semi-supervised, and unsupervised. Supervised methods often use the heuristic that examples belonging to the same class should be close to each other, while those from different classes should be farther apart. Semi-supervised methods use the information in the form of pairwise similarity or dissimilarity constraints. Unsupervised methods learn a distance metric that preserves the geometric relationships (i.e., distance) between most of the training data for the purpose of unsupervised dimensionality reduction. In this thesis, we focus on supervised distance metric learning. The main aim is to develop efficient and scalable algorithms for solving distance metric learning problems under different types of supervision. The proposed algorithms are supported by empirical as well as theoretical studies.
Keywords
Distance metric learning, nearest neighbor, Classification, Machine learning, Optimization, Large-scale learning, Kernel learning

Downloads

  • (...).pdf
    • full text
    • |
    • UGent only
    • |
    • PDF
    • |
    • 7.44 MB

Citation

Please use this url to cite or link to this publication:

Chicago
Nguyen Cong, Bac. 2019. “Supervised Distance Metric Learning for Pattern Recognition”. Ghent, Belgium: Ghent University. Faculty of Bioscience Engineering.
APA
Nguyen Cong, B. (2019). Supervised distance metric learning for pattern recognition. Ghent University. Faculty of Bioscience Engineering, Ghent, Belgium.
Vancouver
1.
Nguyen Cong B. Supervised distance metric learning for pattern recognition. [Ghent, Belgium]: Ghent University. Faculty of Bioscience Engineering; 2019.
MLA
Nguyen Cong, Bac. “Supervised Distance Metric Learning for Pattern Recognition.” 2019 : n. pag. Print.
@phdthesis{8612141,
  abstract     = {Much like in other modeling disciplines does the distance metric used (a measure for dissimilarity) play an important role in the growing field of machine learning. Often, predefined distance metrics (e.g. the Euclidean one) are used to perform such measurement. Unfortunately, most of them ignore any statistical properties that might be estimated from the data. The notion of a good distance metric changes when one moves from one domain to another. For instance, in the problem of computing the dissimilarity for human images, two images could be considered as being similar due to one of the following reasons, the two images are taken from two persons with the same gender, the same age, or the same race. Clearly, it is difficult to use the same distance metric for gender, age, and race since two images might be similar in one case, while being dissimilar in the other case. For this reason, most research efforts have been devoted to automatically learn a good distance metric from data. Depending on the availability of training data, distance metric learning methods can be divided into three categories: supervised, semi-supervised, and unsupervised. Supervised methods often use the heuristic that examples belonging to the same class should be close to each other, while those from different classes should be farther apart. Semi-supervised methods use the information in the form of pairwise similarity or dissimilarity constraints. Unsupervised methods learn a distance metric that preserves the geometric relationships (i.e., distance) between most of the training data for the purpose of unsupervised dimensionality reduction. In this thesis, we focus on supervised distance metric learning. The main aim is to develop efficient and scalable algorithms for solving distance metric learning problems under different types of supervision. The proposed algorithms are supported by empirical as well as theoretical studies.},
  author       = {Nguyen Cong, Bac},
  isbn         = {9789463571968},
  language     = {eng},
  pages        = {XXV, 293},
  publisher    = {Ghent University. Faculty of Bioscience Engineering},
  school       = {Ghent University},
  title        = {Supervised distance metric learning for pattern recognition},
  year         = {2019},
}