η-learning framework predicts rare events without observed extreme data

bravo_abad · x · 2026-08-22

Kai Chang and coauthors introduce the Extreme Event Aware (η-learning) framework to address the difficulty of predicting rare events like floods or structural failures in scientific machine learning. Standard models fail in data-scarce regimes when optimizing for average error. η-learning overcomes this by constraining the statistics of extremeness-associated observables during training, utilizing qualitative scientific knowledge or unlabeled data instead of requiring observed extreme samples in the labeled set.

Original post →

More from Research

Research channel →