Learning With Limited Supervision
Learning with limited supervision encompasses various paradigms, including unsupervised, self-supervised, few-shot, and continual learning. Unsupervised and self-supervised learning address the challenge of leveraging the large availability of unlabeled data to avoid relying on labeled data, which can be time-consuming and expensive to obtain. Few-shot learning aims to acquire knowledge from a small amount of data, particularly in domains like medical applications, where traditional supervised approaches may need more samples. Continual learning focuses on preventing catastrophic forgetting, a phenomenon that occurs when updating model weights for current tasks negatively impacts performance on prior tasks. Our group primarily focuses on developing approaches utilizing these learning paradigms for image classification and segmentation, as well as video applications like panoptic segmentation or action recognition.