Generative Models in Continual Learning, Domain Adaptation, and Image Translation

Presenter: Professor Orcun Goksel,  Dept. of Information Technology, Uppsala University, Sweden

Abstract: In this two-part presentation, I will start by highlighting the importance and value of accumulating incremental knowledge in learned models. I will introduce generative models for image replay and feature alignment in scenarios of domain and class incremental learning. These contributions will be illustrated through applications of image classification and segmentation in natural and medical images, such as digital pathology, fundus photography, and magnetic resonance imaging.

In the second part, I will present our developments towards simulated medical training of sonographers in a virtual reality environment. To that end, we have devised real-time techniques for generating ultrasound images from 3D anatomical models, as well as integrating generative models and label-to-image translation to bridge the gap between simulated and clinical ultrasound appearances. This will be shown to have resulted in one of the most realistic training simulators of its kind to date.

‍

‍