Generalization & Adaptation under Distribution Shift

Developing machine-learning methods that generalize beyond their training distributions and adapt efficiently to new domains, tasks, and environments with limited supervision.

Research spans cross-domain multitask learning (ICML 2012); provable adaptation across multiway domains through representation learning, enabling zero-shot transfer and compositional generalization to unseen domains (ICLR 2022); and transferable reward learning with test-time policy adaptation for generalization to unseen target classes (ICLR 2022).

Recent work extends these ideas to real-world physical sensing systems, including studying the role of inductive biases in generalization and validating the resulting methods through field experiments under different sensing conditions (JLT 2024), as well as adapting pretrained multimodal foundation models to downstream tasks under severe domain shifts with CLAP-S (ICASSP 2025).

Efficient Foundation-Model Adaptation & Post-Training →