UKRI Safe and Trusted AI PhD student, Patrick Henriksen (2019 cohort) presented his paper, ‘Robust Training of Neural Networks against Bias Field Pertubations’ at the 37 th AAAI Conference on Artificial Intelligence 2023 (AAAI 2023).
The performance of neural networks can often be severely affected by slight modifications to the networks’ inputs; i.e. they are not robust to input perturbations. As Patrick explains, “In this paper, we consider methods for training networks such that they are robust to a class of smooth spatially varying intensity perturbations modelled via bias fields. We then empirically show that classification networks trained with our novel approach are more robust to bias field perturbations with only slightly reduced accuracy”.
AAAI 2023 took place in Washington DC and was Patrick’s opportunity to attend a conference in-person for the first time, learn from other researchers and even squeeze in some sightseeing. Patrick said, “It felt great to finally meet many of the people I had only seen in virtual conferences and meetings. Attending and presenting in person also facilitated spontaneous discussion and made it easier to meet other researchers in the field; thus, it felt more rewarding than the virtual conferences we have become used to over the last few years”.
Patrick is in his fourth year of the his PhD and appreciates how being part of the CDT has made his PhD experience more social. He enjoys working with other PhD students in similar fields, and says, “the CDT seminars and activities cover interesting topics and subjects and aid as an inspiration for future research”.
The paper is co-authored by Alessio Lomuscio.