Lu, Xiaoou;
(2021)
Weakly supervised learning with stochastic supervision and knowledge transfer.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Xiaoou_Lu_thesis.pdf Download (1MB) | Preview |
Abstract
In recent years, machine learning methods especially supervised learning methods have achieved great progress in both methodologies and applications. However, in supervised learning, each training sample requires a label to indicate its ground-truth. In many machine learning tasks, it is hard to get sufficient accurately labelled training samples. Weakly supervised learning is an extended setting of supervised learning to more general tasks. In this thesis, we focus on proposing novel methods for inaccurate supervision and incomplete supervision under the setting of weakly supervised learning. In inaccurate supervision, problems with nondeterministic labels, such as stochastic supervision problems, are rarely discussed. In stochastic supervision, the supervision is a probabilistic assessment rather than a deterministic label. In Chapter 2, we provide four generalisations of stochastic supervision models, extending them to asymmetric assessments, multiple classes, feature-dependent assessments, and multi-modal classes, respectively. Corresponding to these generalisations, four new EM algorithms are derived. We show the effectiveness of our generalisations through illustrative examples of simulated datasets, as well as real-world examples of two famous datasets, the MNIST dataset, and the CIFAR-10 dataset. For incomplete supervision problems, we focus on improving the semi-supervised learning in one domain/task by transferring knowledge from another domain/task or from many domains/tasks. In Chapter 3, a novel domain-adaptation-based method is proposed to improve a typical application of semi-supervised learning: the pose estimation, in which the implicit density estimation problem in the domain adaptation is solved by using a neural network to approximate it. The proposed method transfers the knowledge from the training samples in the synthetic data domain to improve the learner in the real data domain, and achieves state-of-the-art performance. In Chapter 4, we focus on transferring knowledge from many tasks to improve the semi-supervised few-shot learning. We use meta-learning to transfer knowledge from many meta-train tasks. A tailor-made ensemble method for few-shot learning is proposed to relieve the pseudo-label noise problem in the semi-supervised few-shot learning. The proposed method also achieves state-of-the-art performances in two widely used benchmark datasets (miniImageNet and tieredImageNet) in few-shot learning.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Weakly supervised learning with stochastic supervision and knowledge transfer |
| Event: | University College London |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2021. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) Licence (https://creativecommons.org/licenses/by-nc-nd/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 Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Statistical Science |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10121185 |
Archive Staff Only
 |
View Item |
