UCLA

This dissertation develops novel semi-supervised and self-supervised learning algorithms, motivated by settings where labeled training data is scarce or unavailable, respectively. We study two core problems in graph-based semi-supervised learning: how to construct a meaningful graph, and how to exploit its structure. To address the former, we derive backpropagation equations to precisely integrate graph construction and graph-based learning into a neural network, so end-to-end training directly learns data embeddings conducive to high-quality graphs. Our novel Graph Learning Layer (GLL) demonstrates improved generalization and adversarial robustness compared to the standard softmax classification head across vision datasets, architectures, and label rates. To tackle the latter, we leverage graph curvature to design coreset selection and active learning routines that dynamically target different regions of the graph according to its topology. We also present an algorithm that modifies the graph's multiscale connectivity based on active learning acquisitions, enhancing local structure. On a variety of benchmark datasets, our methods consistently outperform existing work in coreset quality, sequential and batch active learning, and computational efficiency, reaching higher accuracy faster and with fewer labels. Extending the theme of label-efficiency to disaster response with remote sensing imagery, we design a self-supervised vision transformer trained on near-global, time-series synthetic aperture radar (SAR) data to detect and delineate land disturbances. On three distinct real-world natural disasters, our approach yields accurate delineations and surpasses prior SAR disturbance methods, representing an important step towards operational and label-free disturbance monitoring using modern artificial intelligence tools.