Beyond the Ideal: Multimodal Learning with Imperfect Data Conditions

Multimodal learning combines heterogeneous data sources such as vision, audio, and text to improve perception and decision-making. Despite its empirical success, this thesis ar- gues that multimodal learning is not inherently robust: most existing approaches assume the availability and reliability of all modalities, leading to significant performance degradation when modalities are missing or corrupted. Our work investi- gates how robustness in multimodal learning can be explicitly modeled and enforced under imperfect data conditions. It addresses three interconnected aspects: architectural design for modality integration, learning with missing modalities, and robustness to corrupted modalities. The contributions include unified taxonomies that reveal structural limitations of existing methods, a modality-agnostic learning framework that enables reliable inference under missing modalities, and principled benchmarks and models for evaluating and im- proving robustness under modality corruption. Overall, our work demonstrates that robustness in multimodal learning must be deliberately designed rather than assumed, and pro- vides foundations for building reliable multimodal systems in real-world environments.