What is Out-of-Distribution Detection?#

OOD mechanisms calculate epistemic uncertainty to flag when test inputs diverge from training distributions, prompting models to abstain rather than issue confidently incorrect predictions.

Everyday Analogy#

A seasoned chemist politely declining to answer a question on high-level tax law because it falls outside their professional training.

Practical Applications#

  • Autonomous driving systems recognizing extreme, unprecedented road conditions
  • Medical AI flagging corrupt or unfamiliar imaging artifacts