Cookies

We use cookies to ensure that we give you the best experience on our website. By continuing to browse this repository, you give consent for essential cookies to be used. You can read more about our Privacy and Cookie Policy.


Durham e-Theses
You are in:

Modelling High-Dimensional Data with Likelihood-Based Generative Models

BOND-TAYLOR, SAMUEL,EDWARD (2024) Modelling High-Dimensional Data with Likelihood-Based Generative Models. Doctoral thesis, Durham University.

[img]
Preview
PDF - Accepted Version
Available under License Creative Commons Attribution 3.0 (CC BY).

54Mb

Abstract

Deep generative models are a class of techniques that use neural networks to form a probabilistic model of training data so that new samples from the data distribution can be generated. This type of learning method necessitates greater level of understanding of the underlying data than common supervised learning approaches. As such, generative models have proven to be useful for much more than simply generation, with applications including, but far from limited to, segmentation, semantic correspondence, classification, image translation, representation learning, solving inverse and ill-posed problems, disentanglement, and out of distribution detection.

This thesis explores how to efficiently scale deep generative models to very high dimensional data in order to aid application to cases where fine details are crucial to capture, as well as data such as 3D models and video. A comprehensive review is carried out, from fundamentals to recent state-of-the-art approaches, to understand the properties of different classes of approaches, diagnose what makes scaling difficult, and explore how techniques can be combined to trade off speed, quality, and diversity. Following this, three new generative modelling approaches are introduced, each with a different angle to enable greater scaling. The first approach enables greater scaling by representing samples as continuous functions thereby allowing arbitrary resolution data to be modelled; this is made possible by using latent conditional neural networks to directly map coordinates to content. The second approach compresses data to a highly informative discrete space then models this space with a powerful unconstrained discrete diffusion process, thereby improving sample quality over the first method while allowing faster sampling and better scaling than comparable discrete methods. The final approach extends diffusion models to infinite dimensional spaces, combining the advantages of the first two approaches to allow diverse, high quality samples, at arbitrary resolutions. Based on these findings and current research, the thesis closes by discussing the state of generative models, future research directions, and ethical considerations of the field.

Item Type:Thesis (Doctoral)
Award:Doctor of Philosophy
Faculty and Department:Faculty of Science > Computer Science, Department of
Thesis Date:2024
Copyright:Copyright of this thesis is held by the author
Deposited On:11 Oct 2024 08:50

Social bookmarking: del.icio.usConnoteaBibSonomyCiteULikeFacebookTwitter