Advancing Segmentation and Unsupervised Learning Within the Field of Deep Learning
Michael Kampffmeyer · NORA - Norwegian Open Research Archives · 2018
Due to the large improvements that deep learning based models have brought to a variety of tasks, they have in recent years received large amounts of attention. However, these improvements are to a large extent achieved in supervised settings, where labels are available, and initially focused on traditional computer vision tasks such as visual object recognition. Specific application domains that consider images of large size and multi-modal images, as well as applications where labeled training data is challenging to obtain, has instead received less attention. This thesis aims to fill these gaps from two overall perspectives. First, we advance segmentation approaches specifically targeted towards the applications of remote sensing and medical imaging. Second, inspired by the lack of labeled data in many high-impact domains, such as medical imaging, we advance four unsupervised deep learning tasks: domain adaptation, clustering, representation learning, and zero-shot learning. The works on segmentation address the challenges of class-imbalance, missing data-modalities and the modeling of uncertainty in remote sensing. Founded on the idea of pixel-connectivity, we further propose a novel approach to saliency segmentation, a common pre-processing task. We illustrate that phrasing the problem as a connectivity prediction problem, allows us to achieve good performance while keeping the model simple. Finally, connecting our work on segmentation and unsupervised deep learning, we propose an approach to unsupervised domain adaptation in a segmentation setting in the medical domain. Besides unsupervised domain adaptation, we further propose a novel approach to clustering based on integrating ideas from kernel methods and information theoretic learning achieving promising results. Based on our intuition that meaningful representations should incorporate similarities between data points, we further propose a kernelized autoencoder. Finally, we address the task of zero-shot learning based on improving knowledge propagation in graph convolutional neural networks, achieving state-of-the-art performance on the 21K class ImageNet dataset.