SQUIDS Seminar - How mathematics can inspire novel deep learning architectures: two case studies
| Dates: | 15 November 2023 |
|---|
| Times: | 14:30 - 16:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Department of Mathematics |
|---|
| Who is it for: | University staff, External researchers, Current University students |
|---|
| Speaker: | Jeremy Budd |
|---|
|
Schedule:
- 2:30pm: pretalk
- 3:05pm: research talk 'How mathematics can inspire novel deep learning architectures: two case studies'
Abstract: Deep learning is revolutionising the modern world, including applied mathematics. Yet, it has infamously been described as "alchemy", and one place that alchemy can arise is in the choice of architecture. Can we leverage mathematical insights to design deep learning architectures in a more principled way?
In this talk, I will give two examples of this from my own research. In the first, I will talk about my recent preprint with Carola-Bibiane Schönlieb, Subhadip Mukherjee, and Zakhar Shumaylov on learned regularisation for inverse problems in imaging; how we were inspired by (non)convex analysis to design a novel input weakly convex neural network, and used this to learn an adversarial convex-nonconvex regulariser with provable guarantees and which overcame the numerical issues of previous adversarial regularisers. In the second, I will talk about ongoing research with Martin Benning, Jonas Latz, Lisa Kreusser, and James Rowbottom, on designing a deep learning architecture for image segmentation based on the graph Merriman-Bence-Osher segmentation method. This method is particularly promising because the graph this method learns on the inputted image provides some built-in explainability of the outputted segmentation.
Speaker
Jeremy Budd
Organisation: Caltech
Travel and Contact Information
Find event
G.114
Alan Turing Building
Manchester
