Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.790446
Title: Beyond linear similarity function learning
Author: Bohné, J. W. N.
ISNI:       0000 0004 8497 9710
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Being able to measure the similarity between two patterns is an underlying task in many machine learning and data mining applications. However, handcrafting an effective similarity function for a specific application is difficult and tedious. This observation has led to the emergence of the topic of similarity function learning in the machine learning community. It consists in designing algorithms that automatically learn a similarity function from a set of labeled data. In this thesis, we explore advanced similarity function concepts: local metric, deep metric learning and computing similarity with data uncertainty. Linear metric learning is a widely used methodology to learn a similarity function from a set of similar/dissimilar example pairs. Using a single linear metric may be a too restrictive assumption when handling heterogeneous datasets. Lately, local metric learning methods have been introduced to overcome this limitation. However, most methods are subject to constraints preventing their usage in many applications. For example, some require the knowledge of all possible class labels during training. In this thesis, we present a novel local metric learning method, which overcomes some limitations of previous approaches. Deep learning has become a major topic in machine learning. Over the last few years, it has been successfully applied to various machine learning tasks such as classification or regression. In this thesis, we illustrate how neural networks can be used to learn similarity functions which surpass linear and local metric learning methods. Often, similarity functions have to deal with noisy feature vectors. In this context, standard similarity learning methods may result in unsatisfactory performance. In this thesis, we propose a method which leverages additional information on the noise magnitude to outperform standard methods.
Supervisor: Pontil, M. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.790446  DOI: Not available
Share: