Generalization & Adaptation under Distribution Shift
How can models remain effective when domains, tasks, and environments differ from their training conditions?
My research connects machine-learning methodology with real-world systems, spanning generalization, efficient adaptation, multimodal learning, and physical sensing.
How can models remain effective when domains, tasks, and environments differ from their training conditions?
How can knowledge from pretrained and specialized models be adapted, compressed, and consolidated efficiently and reliably?
How far can small multimodal models progress from open-vocabulary recognition toward understanding and reasoning?
Machine learning for optical fiber sensing over deployed telecommunication networks.