UCL Discovery
UCL home » Library Services » Electronic resources » UCL Discovery

Scaling Machine Learning Systems using Domain Adaptation

Mathur, Akhil; (2020) Scaling Machine Learning Systems using Domain Adaptation. Doctoral thesis (Ph.D), UCL (University College London).

[img] Text
PhD_Thesis_Akhil Mathur.pdf - Accepted version
Access restricted to UCL open access staff until 1 July 2021.

Download (21MB)


Machine-learned components, particularly those trained using deep learning methods, are becoming integral parts of modern intelligent systems, with applications including computer vision, speech processing, natural language processing and human activity recognition. As these machine learning (ML) systems scale to real-world settings, they will encounter scenarios where the distribution of the data in the real-world (i.e., the target domain) is different from the data on which they were trained (i.e., the source domain). This phenomenon, known as domain shift, can significantly degrade the performance of ML systems in new deployment scenarios. In this thesis, we study the impact of domain shift caused by variations in system hardware, software and user preferences on the performance of ML systems. After quantifying the performance degradation of ML models in target domains due to the various types of domain shift, we propose unsupervised domain adaptation (uDA) algorithms that leverage unlabeled data collected in the target domain to improve the performance of the ML model. At its core, this thesis argues for the need to develop uDA solutions while adhering to practical scenarios in which ML systems will scale. More specifically, we consider four scenarios: (i) opaque ML systems, wherein parameters of the source prediction model are not made accessible in the target domain, (ii) transparent ML systems, wherein source model parameters are accessible and can be modified in the target domain, (iii) ML systems where source and target domains do not have identical label spaces, and (iv) distributed ML systems, wherein the source and target domains are geographically distributed, their datasets are private and cannot be exchanged using adaptation. We study the unique challenges and constraints of each scenario and propose novel uDA algorithms that outperform state-of-the-art baselines.

Type: Thesis (Doctoral)
Qualification: Ph.D
Title: Scaling Machine Learning Systems using Domain Adaptation
Event: UCL (University College London)
Language: English
Additional information: Copyright © The Author 2020. Original content in this thesis is licensed under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) Licence (https://creativecommons.org/licenses/by/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request.
UCL classification: UCL
UCL > Provost and Vice Provost Offices
UCL > Provost and Vice Provost Offices > UCL BEAMS
UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science
UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science
URI: https://discovery.ucl.ac.uk/id/eprint/10117130
Downloads since deposit
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item