Open-World Learning: From limited observations to an unbounded world

PhD Qualifying Examination


Title: "Open-World Learning: From limited observations to an unbounded world"

by

Miss Zhenqi HE


Abstract:

Modern machine learning relies on finite training observations, yet the 
concepts, compositions, and distributions encountered in deployment extend 
beyond their coverage. Open-world learning therefore requires learning and 
reasoning beyond what has been directly observed. Existing research addresses 
this challenge through separate tasks for unseen concepts, compositions, and 
domains, leaving their shared observation gaps less explicit. This survey 
provides a unified perspective grounded in open-world learning. We first 
taxonomize limited observations into limited semantic observations, including 
limited primitives and compositions, and limited distributional observations, 
arising from limited domains. We then examine two complementary strategies: 
transfer beyond observations, which reuses existing knowledge to generalize 
beyond direct supervision, and discovery beyond observations, which identifies 
unknowns and uncovers new structure. We offer a comprehensive survey of these 
paradigms across static and evolving settings, their limitations, and the 
opportunities and challenges introduced by foundation models and multimodal 
large language models (MLLMs), highlighting knowledge boundaries and future 
directions toward truly open-world learning.


Date:                   Monday, 26 October 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Dr. Long Chen (Supervisor)
                        Dr. Hao Chen (Chairperson)
                        Dr. Dan Xu