Machine-learning models are usually developed under an implicit assumption: future inputs will resemble the data used for training and evaluation. When operating conditions change, that assumption can fail while a model continues to produce confident predictions.
My work in this area studies out-of-distribution and anomaly detection as forms of model monitoring. It includes representation-learning methods, outlier exposure, and detectors augmented with knowledge about the structure of the expected data. Across these approaches, the objective is to produce useful evidence that a learned model may be operating beyond its competence.
