Hayes, Peter;
(2024)
Towards More Data Efficient Deep Learning.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Hayes__Thesis.pdf - Accepted Version Download (3MB) | Preview |
Abstract
Deep learning is the most predominant learning paradigm in artificial intelligence. The deep neural network models deployed in practise are increasingly data and resource hungry. This thesis introduces several methodological ideas to improve the data efficiency of deep learning algorithms across a diverse range of applications. The first section deals with supervised deep learning in settings where collecting labelled data is expensive in time and/or cost. We focus on the scenario where multiple weak and relatively cheap sources of supervision are also available. We develop an approach that jointly trains the supervised model and a separate label model to aggregate weak supervision sources and show it outperforms existing weak learning approaches across a benchmark of natural language processing problems. The second section focuses on unsupervised deep learning; specifically the problem of generative modelling. We study generalisation in variational inference when neural network based amortization is used. We introduce a wake-sleep style training scheme for variational autoencoders that improves generalization performance for a given budget of training data and demonstrate the utility of this approach in image modelling and compression applications. The third section explores how to improve the efficiency of deep reinforcement learning (RL). We propose a model-based RL framework that learns a low dimensional representation of the environment while avoiding the need to learn a generative model of the environment. We demonstrate gains in efficiency over model-free methods when learning directly from pixels in a control problem. The final section tackles how to align large pretrained generative models to human preferences. We discuss an alternative approach to reinforcement learning from human feedback based on a maximum likelihood criterion and introduce a simple active learning regime for more efficiently collecting preference data.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Towards More Data Efficient Deep Learning |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/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 BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10199689 |
Loading...
Loading...Loading...Loading...Archive Staff Only
 |
View Item |
