Supervisor: Dr Adrian Baule
Project description:
Generative artificial intelligence (AI) models have recently driven remarkable technological advances in image synthesis and natural language processing with the potential to revolutionize not only relevant industrial sectors but societal interactions as a whole. At the forefront of this progress are diffusion models, which provide powerful tools to generate new samples from complex underlying data distributions represented, e.g., by realistic images. While a lot of work has been devoted to improving the empirical performance of such models in view of specific applications, a comprehensive theoretical understanding has still proven challenging. Recent work has established surprising connections between diffusion models and the statistical physics of disordered stochastic many-particle systems exhibiting phase transitions and spontaneous symmetry breaking. These approaches open up novel methodology to provide a deeper understanding of the dynamics of diffusion models that could provide important steps towards better explainable and controllable generative AI.
The aim of this PhD project is to investigate a novel class of score-based diffusion models driven by a combination of Gaussian and non-Gaussian noises representing the effect of jumps in the generative process. The main focus of the work will be on theoretical approaches based on statistical physics methods to understand the metastable dynamics of such jump-diffusion generative models. Both analytically tractable simplified data distributions and realistic high-dimensional image data from known databases will be analysed. The project will implement computational sampling algorithms, investigate and optimize their efficiency, and compare their model performance with that of conventional diffusion models without jumps. A particular question addressed by the project will be whether the inclusion of jumps can prevent the empirically observed diversity reduction by many AI image synthesizers, which implement a bias, e.g., towards a majority ethnic group.
The project will be embedded in the world-leading research activities at the QMUL Centers for "Complex Systems" and "Probability, Statistics, and Data Science". The prospective PhD student will also benefit from numerous opportunities to participate in further research, training, and networking activities in the wider London area.
Further information:
How to apply Entry requirements Fees and funding