Gradient methods, such as gradient descent and stochastic gradient descent, achieve remarkable performances in neural network training but suffer typically from strong instability that makes the optimization of specific architectures challenging, time-consuming, and susceptible to attacks. For this reason, a dynamic regularization of the optimization algorithm is often necessary. Our goal is to lay a solid theoretical foundation able to unify different regularization strategies by looking at the gradient-flow structure of the training algorithm. We will use these theoretical insights to study sparsity properties of neural networks during training, analyzing how they depend on the chosen regularization. We will then develop a robustness theory for dynamically regularized neural networks able to explain and defend against adversarial attacks. Applications to biological data-driven models such as CT-reconstruction, single-particle tracking (SPT) for fluorescence microscopy and microbubbles flow for drug delivery will be considered.
The PhD candidate will work under the supervision of Dr. Marcello Carioni and will be part of the group Mathematics of Imaging and Artificial Intelligence (MIA) headed by Prof. Christoph Brune at the department of Applied Mathematics. There will be plenty of opportunities for collaborations with researchers in group of Prof. Carola Schönlieb at the University of Cambridge and in the group of Prof. Kristian Bredies at KFU Graz.
Information and application
Are you interested in this position? Please submit your application before July 10, 2022 via the ‘Apply now’ button below and include:
- A motivation letter, emphasizing your specific interest and motivation to apply for a PhD position in this research area.
- A detailed Curriculum Vitae.
- An academic transcript of BSc and MSc education, including grades.
- A short description of your MSc thesis/final project.
- References (contact information) of two scientific staff members (one of whom should be the supervisor of your MSc thesis/final project) who are willing to provide a recommendation letter at our request.
We particularly encourage/support female applicants to apply.
For more information regarding this position, you are welcome to contact Dr. Marcello Carioni email@example.com
About the organization
The faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS) uses mathematics, electronics and computer technology to contribute to the development of Information and Communication Technology (ICT). With ICT present in almost every device and product we use nowadays, we embrace our role as contributors to a broad range of societal activities and as pioneers of tomorrow's digital society. As part of a people-first tech university that aims to shape society, individuals and connections, our faculty works together intensively with industrial partners and researchers in the Netherlands and abroad, and conducts extensive research for external commissioning parties and funders. Our research has a high profile both in the Netherlands and internationally. It has been accommodated in three multidisciplinary UT research institutes: Mesa+ Institute, TechMed Centre and Digital Society Institute.