Word
out of distribution
Definition
adj. describing new data or situations that are completely different from anything an AI model saw during its training.
adj. originating from a probability distribution distinct from that of the training dataset, posing severe generalization and safety risks for machine learning systems.
Examples
“The image classifier made a strange mistake because the blurred photograph was completely out-of-distribution.”
“Reliable safety systems must detect out-of-distribution inputs and alert human supervisors rather than guessing wildly.”
“Energy-based models provide robust out-of-distribution detection by mapping anomalous sample vectors to substantially higher free energy scores.”
Examples
simple
“The image classifier made a strange mistake because the blurred photograph was completely out-of-distribution.”
contextual
“Reliable safety systems must detect out-of-distribution inputs and alert human supervisors rather than guessing wildly.”
complex
“Energy-based models provide robust out-of-distribution detection by mapping anomalous sample vectors to substantially higher free energy scores.”
Real-World Examples
“Successful deployment of machine learning systems requires that the system be able to distinguish between data that is anomalous or significantly different from that used in training. This is particularly important for deep neural network classifiers, which might classify such out-of-distribution (OOD) inputs into in-distribution classes with high confidence.” “A model that is consistently overconfident will appear to have low perplexity on the training distribution but will show sudden large spikes in perplexity on out-of-distribution text because its over-fitted probabilities do not generalize.” Etymology
Hyphenated compound adjective in machine learning and statistics, contrasting with in-distribution; formalized in empirical risk minimization.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/out-of-distribution