Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form

Efficiently and Effectively Learning Models of Similarity from Human Feedback

Heim, Eric (2016) Efficiently and Effectively Learning Models of Similarity from Human Feedback. Doctoral Dissertation, University of Pittsburgh. (Unpublished)

[img]
Preview
PDF (Final Submission)
Primary Text

Download (5MB)

Abstract

Vital to the success of many machine learning tasks is the ability to reason about how objects relate. For this, machine learning methods utilize a model of similarity that describes how objects are to be compared. While traditional methods commonly compare objects as feature vectors by standard measures such as the Euclidean distance or cosine similarity, other models of similarity can be used that include auxiliary information outside of that which is conveyed through features. To build such models, information must be given about object relationships that is beneficial to the task being considered. In many tasks, such as object recognition, ranking, product recommendation, and data visualization, a model based on human perception can lead to high performance. Other tasks require models that reflect certain domain expertise. In both cases, humans are able to provide information that can be used to build useful models of similarity. It is this reason that motivates similarity-learning methods that use human feedback to guide the construction of models of similarity.

Associated with the task of learning similarity from human feedback are many practical challenges that must be considered. In this dissertation we explicitly define these challenges as being those of efficiency and effectiveness. Efficiency deals with both making the most of obtained feedback, as well as, reducing the computational run time of the learning algorithms themselves. Effectiveness concerns itself with producing models that accurately reflect the given feedback, but also with ensuring the queries posed to humans are those they can answer easily and without errors. After defining these challenges, we create novel learning methods that explicitly focus on one or more of these challenges as a means to improve on the state-of-the-art in similarity-learning. Specifically, we develop methods for learning models of perceptual similarity, as well as models that reflect domain expertise. In doing so, we enable similarity-learning methods to be practically applied in more real-world problem settings.


Share

Citation/Export:
Social Networking:
Share |

Details

Item Type: University of Pittsburgh ETD
Status: Unpublished
Creators/Authors:
CreatorsEmailPitt UsernameORCID
Heim, Ericeric@cs.pitt.eduETH13
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairHauskrecht, Milosmilos@cs.pitt.eduMILOS
Committee MemberHwa, Rebeccahwa@cs.pitt.eduREH23
Committee MemberWang, Jingtaojingtaow@cs.pitt.eduJINGTAOW
Committee MemberSingh, Aartiaarti@cs.cmu.edu
Date: 19 January 2016
Date Type: Publication
Defense Date: 20 November 2015
Approval Date: 19 January 2016
Submission Date: 16 November 2015
Access Restriction: No restriction; Release the ETD for access worldwide immediately.
Number of Pages: 154
Institution: University of Pittsburgh
Schools and Programs: Dietrich School of Arts and Sciences > Computer Science
Degree: PhD - Doctor of Philosophy
Thesis Type: Doctoral Dissertation
Refereed: Yes
Uncontrolled Keywords: Similarity Learning, Kernel Learning, Metric Learning, Online Learning, Active Learning.
Date Deposited: 19 Jan 2016 20:40
Last Modified: 19 Dec 2016 14:42
URI: http://d-scholarship.pitt.edu/id/eprint/26344

Metrics

Monthly Views for the past 3 years

Plum Analytics


Actions (login required)

View Item View Item