The course will deal with advanced deep learning techniques used to solve critical inference problems in remote sensing, image analysis, and computer vision. We plan to cover the following learning paradigms which consider learning a classification model under different levels of supervision. Meta- learning and few-shot learning Zero-shot learning Self-supervised learning Multi-modal learning Multi-task learning Continual learning Domain adaptation and generalization Uncertainty estimation in deep learning We will consider the application of the aforementioned learning algorithms for visual inference tasks with focus to image classification and semantic segmentation. To this end, land-cover classification from remote sensing images of varied modalities would be the running application example that will be discussed at length for all the algorithms. In addition, we will consider a few challenging classification problems in remote sensing (cross-sensor data retrieval, change detection, and land cover map generation for geographically disjoint and large areas) and show how some of these learning techniques (domain adaptation, uncertainty estimate) can be modified to solve these problems. Finally, annotation is a major problem in remote sensing, we plan to discuss how off-the-shelves self-supervision can be used as an alternative in this regard. Like the examples from the remote sensing area, we will consider examples from natural image and video analysis literature concerning very large-scale datasets (ImageNet, DomainNet, to name a few).

• Text / References:
• Goodfellow, Ian, Yoshua Bengio, and Aaron Courville.302240Deep learning. MIT press, 2016.Research papers from IEEE T. Geoscience & Remote Sensing, IEEE T. PAMI, IJCV, CVPR, ICCV, ECCV, NeurIPS, ICML, BMVC, ICLR.Deep learning lectures from Stanford university, NYU (available on youtube), NPTEL. Cliff Greve (Ed.), 'Digital Photogrammetry - An addendum to the Manual of Photogrammerty', ASPRS,1996.