UCLA

The rapid evolution of Artificial Intelligence (AI) has significantly transformed various sectors, particularly in processing large-scale datasets. Traditional supervised learning approaches, which rely on extensive human-annotated datasets, often fall short in capturing the nuanced and hierarchical nature of real-world data, especially in fields like medical diagnostics where expert annotation is indispensable. Consequently, there is a pressing need for automated, scalable methods to discover and model data modalities.

Leveraging deep learning, this dissertation addresses the critical challenges of modality discovery and modeling within large-scale heterogeneous data using advanced AI techniques, focusing on medical applications and 3D motion data generation. This dissertation employs both weakly supervised and self-supervised learning to automate the discovery of modalities within complex datasets. It introduces innovative methods for identifying pathological and physiological biomarkers in intracranial and scalp electroencephalograms (EEG), significantly advancing the treatment of epilepsy, and explores self-supervised learning techniques for understanding and generating 3D motion sequences, crucial for applications in animation, virtual reality, and human-computer interaction. Key contributions include weakly and self-supervised approaches for discovering biomarkers in intracranially and scalp-recorded EEG, a software platform for deploying these detectors in clinical settings, and self-supervised methods for generating diverse 3D motion sequences. This work significantly enhances AI's ability to interpret and generate complex data, with substantial implications for medical diagnostics and motion data applications. The dissertation concludes with a summary of contributions and a discussion of potential future research directions, highlighting the continued need for advancements in automated modality discovery and modeling in AI.