My research aims to pave the way to a reliable Open-world Machine Learning system, which revolves around three aspects:

Out-of-distribution (OOD) Detection: In the open world, the AI system will encounter new contexts and data that were not taught to the algorithms during training. A reliable machine learning model should not only accurately classify in-distribution (ID) samples but also possess the capability to identify samples that lie outside the known distribution.
Open-world Representation Learning (ORL): Moving beyond OOD detection, the ORL problem further requires models to learn the hidden classes within OOD samples, in addition to the known classes.
Interpretability: We aim to build an interpretable machine learning system, particularly in visualizing and quantitating the model’s inference mechanism or even constructing one inherently sparse, simulatable, modular, and thus, interpretable.