Definition and Measurement of Information Risk in OWML

Define and rigorously measure information risk in open-world machine learning by formalizing the quantifiable exposure of models to unknown or novel information sources under non-stationary and uncertain conditions.

Background

Traditional risk formulations focus on error under fixed distributions, but OWML requires accounting for risks arising from novelty, distribution shifts, and uncertainty. The paper argues for an "Open-space Information Risk" perspective that balances known, unknown, and novel informational components.

A formal definition and measurement methodology for information risk would enable provable bounds and safety criteria in open-world systems, complementing existing notions like open-space risk and PAC-Bayes-based generalization by explicitly quantifying informational exposure.