Towards Adaptive and Robust Learning from Multi-Source Data

This thesis aims to address the challenges of adaptivity and robustness in machine learning on multi-source data. Specifically, it investigates three core problems: (1) conflict resolution in multiview learning, which focuses on reconciling inconsistencies across different views of the same data; (2) multimodal recognition with missing modalities, which examines how to maintain performance when one or more input modalities are absent; and (3) logits compression for data-efficient fine-tuning of multimodal large language models (MLLMs), which targets the computational bottlenecks of adapting billion-parameter models. Through addressing these challenges, the thesis contributes three novel methods to the machine learning community and paves the way toward more adaptive and robust learning from diverse data sources, as reflected in the title.