In overall, I am interested in learning with imperfect data (e.g., limited, noisy, imbalanced data), uncertainty estimation and adaptation (continual learning, transfer learning, out-of-distribution detection, domain adaptation), and black-box optimization (Bayesian optimization). My long-term research plan is to develop machine-learning algorithms that can learn continuously with minimal data or supervision and effectively adapt to other domains outside of training distribution, towards efficient open-world machine-learning.
In particular, I am interested in exploring both meta-learning and domain adaptation and integrating them in joint settings on computer vision tasks, i.e. continual few-shot object detection. However, each research field has its own limitations, and combining them often results in extremely complex and constrained settings. Therefore, I am aiming to approach and formulate these joint settings from an optimization perspective, especially in the black-box context (derivative-free), to find effective solutions for solving them and then advance toward efficient algorithms for open-world machine learning.
In addition, by effectively optimizing black-box function(s) under various constraints and conditions based on meta-learning and domain adaptation, such as few-shot evaluations, conflicting objectives, or function shifts, we can develop flexible and robust optimization algorithms for a wide range of fundamental and applied scientific research, such as experimental design for material construction or self-driving car.