Pioneer Centre for AI Lunch Talk: Robert Jenssen

Title

Learning from limited data - few shot learning for medical image segmentation and prototype-based Explainability

Abstract

This talk addresses elements of the challenge of learning from limited data. In particular, the talk puts focus on a prototype-based approach both for few-shot learning and for self-explainable deep learning models, developed at the Visual Intelligence research centre. For few-shot learning, a new anomaly detection-inspired method for image segmentation is presented where self-supervised learning is a key element and where prototypes are designed to focus on foreground objects and not on background objects. Furthermore, as a common mitigation to the limited data problem is to incorporate data from several sources, we investigate self-explainable deep learning to expose label bias from these sources and to expose underlying artifacts that may cause such bias.   

Robert Jenssen is Professor and Head of the Machine Learning Group and leader of the Centre for Visual Intelligence at UiT in Norway, and he is a part-time professor at DIKU, in the Pioneer Centre for AI.