As machine learning systems are deployed in increasingly open-ended and high-stakes environments, their fundamental limitations become more apparent. Distribution shift, adversarial manipulation, catastrophic forgetting, hallucinations, safety failures, and misalignment all stem from a common challenge: making decisions at the boundary between what a model knows and what it does not know. Yet most contemporary learning systems are designed to produce precise predictions even when the available information is incomplete or unreliable.

 

In this talk, I will present recent work on robust out-of-distribution generalisation and imprecise forecasting that seeks to address this challenge. The unifying theme is learning under partial knowledge: developing intelligent systems that can explicitly represent, reason about, and communicate the limits of their own knowledge. Rather than treating ignorance as a hidden vulnerability, these approaches elevate it to a first-class object of inference, enabling models to make more robust predictions and more trustworthy decisions under uncertainty.