Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.819933
Title: Essays on distance metric learning
Author: Yang, Xiaochen
ISNI:       0000 0004 9359 9154
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Many machine learning methods, such as the k-nearest neighbours algorithm, heavily depend on the distance measure between data points. As each task has its own notion of distance, distance metric learning has been proposed. It learns a distance metric to assign a small distance to semantically similar instances and a large distance to dissimilar instances by formulating an optimisation problem. While many loss functions and regularisation terms have been proposed to improve the discrimination and generalisation ability of the learned metric, the metric may be sensitive to a small perturbation in the input space. Moreover, these methods implicitly assume that features are numerical variables and labels are deterministic. However, categorical variables and probabilistic labels are common in real-world applications. This thesis develops three metric learning methods to enhance robustness against input perturbation and applicability for categorical variables and probabilistic labels. In Chapter 3, I identify that many existing methods maximise a margin in the feature space and such margin is insufficient to withstand perturbation in the input space. To address this issue, a new loss function is designed to penalise the input-space margin for being small and hence improve the robustness of the learned metric. In Chapter 4, I propose a metric learning method for categorical data. Classifying categorical data is difficult due to high feature ambiguity, and to this end, the technique of adversarial training is employed. Moreover, the generalisation bound of the proposed method is established, which informs the choice of the regularisation term. In Chapter 5, I adapt a classical probabilistic approach for metric learning to utilise information on probabilistic labels. The loss function is modified for training stability, and new evaluation criteria are suggested to assess the effectiveness of different methods. At the end of this thesis, two publications on hyperspectral target detection are appended as additional work during my PhD.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.819933  DOI: Not available
Share: