Michele LaurelliBlog

Out-of-distribution (OOD)

Out-of-distribution (OOD): A condition where an input differs sufficiently from training data to make prediction unreliable. OOD detection uses confidence scores, distances or auxiliary models, but no universal threshold exists. A robust system can abstain or route the case to human review.

In italiano: Out-of-Distribution (OOD)Legge V · Overfitting, bias e collasso dei modelli

OOD detection uses confidence scores, distances or auxiliary models, but no universal threshold exists. A robust system can abstain or route the case to human review.

Examples

  • Out-of-domain image for an industrial CNN
  • Language never seen by the classifier
  • Sensor with an anomalous signal