Out-of-distribution Data

This is a persistent and critical production issue present in any machine learning and deep learning systems, no matter how good the models were trained.

A single out-of-distribution sample data fed into a well-trained model can result in a few problematic outcomes, and in production-level systems they become fat-tailed critical risks that have outsized consequences.

Problematic consequences of out-of-distribution data

1. False positive & False Negative: prediction does not match the ground truth.
2. Unreliable True positive & True Negative: prediction matches the ground truth but due to samples being out-of-distribution, the performance of these predictions becomes very erratic.