Uncertainty Modelling for Open Domain Generalization

Open domain generalization asks a model to perform on domains absent from training, where even the set of possible labels may differ. This thesis studies how uncertainty should be represented and used in that setting, so a model can recognize when it is operating outside what it knows.

Overview

Problem — standard generalization assumes shared label spaces and related domains; open domain generalization drops both, making confident-but-wrong predictions the core failure mode.
Approach — characterize entropy-, dropout-, and ensemble-based uncertainty across distribution shifts, and design an uncertainty-based decision framework for open-set prediction.
See the Research page for the full thesis entry.