In many fields high-dimensional inverse imaging problems are encountered. For example, imaging the raw data acquired by radio interferometric (RI) telescopes involves solving an ill-posed inverse problem to recover an image of the sky from noisy and incomplete Fourier measurements. Future telescopes, such as the Square Kilometre Array (SKA), will usher in a new big-data era for radio interferometry, with data rates comparable to world-wide internet traffic today. The recently proposed SPIDER optical interferometric sensor could revolutionalise space observations, due to low weight, low power consumption, and high resolution, but again involves solving an ill-posed inverse problem. The SPIDER concept could also be useful for sensing the surrounding world, e.g. with panels built into autonomous vehicles. Magnetic resonance (MR) imaging (MRI) involves solving a very similar inverse problem to interferometric imaging. MRI also encounters a growing big-data problem as new imaging modalities are considered (e.g. 3D and diffusion MRI) and are pushed to higher resolutions.
Artificial intelligence (AI) techniques are typically applied to regression and classification problems. While imaging can be considered a type of regression problem, imaging on the whole presents a different type of modality than standard regression problems. Applying machine learning to imaging, rather than standard classification or regression, is a relatively new and immature field, although much progress has been made in recent years. In general when learning to image, “fully-learned” approaches that do not exploit any knowledge of the physics of the problem are difficult since the parameterisation of arbitrary inverse operators must inevitably be very high dimensional. Instead we will take a “knowledge- driven” approach, exploiting the physics of the problem that is encoded in the measurement operator in a deep neural network setting. We will consider network architectures consisting of convolutional and deconvolutional layers. We will develop stochastic models of permissible measurement operators for training to address the issue that the measurement operator may changes with each observation. In additional, we will explore the use of “learnt iterative” schemes that more tightly integrate the physics of the problem with the learning algorithm. We will also investigate the inclusion of uncertainty quantification (e.g. error estimation) into learning to image frameworks, which is an important missing component in many imaging problems, extending recent developments by McEwen to the deep learning setting.
Applications submitted by 31st January 2020 will be given full consideration. We will continue accepting applications until all places are filled. After we receive your application, we will select candidates for interviews. If you are selected, you will be invited for an interview at MSSL. You will have the opportunity to see the laboratory, students' flats and talk to current students. The studentships are for the advertised projects only. In your application, please specify which project you want to apply for.
To apply, please visit the Online Application page, select department of "Space & Climate Physics" and programme type of "Postgraduate Research". After pushing "Search Now" button, select "RRDSPSSING01: Research Degree: Space and Climate Physics" for Full-time or Part-time mode.
Our Online Applications page can be found here: https://www.ucl.ac.uk/adminsys/search/
An upper second-class Bachelor’s degree, or a second-class Bachelor’s degree together with a Master's degree from a UK university in a relevant subject, or an equivalent overseas qualification.
Students from the UK or those from the EU who meet the residency requirements (3 years' full-time residency in the UK) are potentially eligible for a Science and Technology Facilities Council (STFC) studentship.