Provable Theoretical Framework for Open-World Machine Learning

Establish a generalizable and provable theoretical framework for open-world machine learning (OWML) that formalizes learning under uncertainty and evolving label spaces and yields rigorous guarantees for system behavior in open environments.

Background

The paper emphasizes that existing OWML methods often rely on heuristics (e.g., confidence thresholds and memory replay) and lack a unified mathematical foundation that explains when models can safely recognize known classes, reject unknowns, and adapt over time. Without such a framework, it is difficult to quantify adaptability limits or provide stability guarantees under nonstationary conditions.

Information theory is proposed as a promising backbone to unify OWML tasks via entropy, mutual information, and divergence, but a generalizable and provable framework that integrates these elements to support formal guarantees across open-world scenarios has not yet been established.