New PhD Vacancy: Theory of Score-based Generative Models

News
CAMERA New PhD Vacancy: Theory of Score-based Generative Models

Funded PhD Vacancy (UK Students Only): Theory of Score-based Generative Models

University of Bath    Department of Computer Science

About the Project

The University of Bath is inviting applications for a PhD project on the Theory of Score-base Generative Models in the Department of Computer Science commencing 30 September 2024.

The successful student will be part of the Centre for the Analysis of Motion, Entertainment Research and Applications (CAMERA) which performs world-leading multi-disciplinary research in Intelligent Visual and Interactive Technology. Funded by the EPSRC and the University of Bath, CAMERA exists to accelerate the impact of fundamental research being undertaken at the University in the Departments of Computer Science, Health and Psychology. The successful candidate will work closely work with the experts from CAMERA and with collaborators from the University of Bristol and project partners associated with the MyWorld programme. The ambitious MyWorld project is funded by the UKRI Strength in Places fund bringing together 30 partners from Bristol and Bath’s creative technologies sector and world-leading academic institutions to create a unique cross-sector consortium.

Overview of this Project:

Deep generative models have been applied across a wide variety of domains including image, shape, text and speech synthesis tasks, and are showing promise for molecule design in material science and drug discovery. Common approaches have included flow-based (Normalizing Flows), latent variable (VAEs) and implicit generative (GANs) models.

In contemporary generative modelling research, solving transport problems, i.e. learning to bridge arbitrary densities, using dynamical methods defined by ordinary or stochastic differential equations (ODEs/SDEs) has attracted significant interest.

A major milestone was the introduction of denoising diffusion models (DDMs). Inspired by non-equilibrium thermodynamics, a forward diffusion perturbs some input data towards random noise while a deep model is tasked with learning to reverse the process. As a result, it can be used to generate desired data samples starting from random noise. This simple idea has quickly led to state-of-the-art performance on a range of downstream tasks, most famously text-to-image generation via the likes of DALLE-3 and Stable Diffusion models.

More recent works have illustrated the potential for Score-based generative models (SBGMs), including diffusion models, to form a powerful class of models which can represent complex distributions over high-dimensional spaces as well as operate directly in function space.

While there have been notable achievements, substantial limitations persist. Specifically, training diffusion models data-intensive and computationally demanding, limiting their practical use for real-world tasks that involve high-dimensional samples from noisy and sparse datasets.

What are you going to do?

Research in this domain is still in its formative stages with considerable potential. Therefore, we are seeking a PhD candidate to drive forward theoretical advancements in this field.

The flexibility in the core formulations provides a rich seam of possibilities for exploring generalizations that retain the robustness of the core mathematical foundations while formulating models which are more adaptable and efficient. For instance, it has recently been shown that DDMs are not strongly dependent on the choice of image degradation, which has opened the possibility for an entire family of generative models can be constructed by varying this choice.

Other key research directions include deriving new formulations to make the latent space interpretable and control of intermediate trajectories, which holds the key to learning models with controllable sample generation at inference time. For further details, please see https://www.ndfcampbell.org/opportunities/

Keep Up To Date?

Join Our Centre Newsletter