Decoder. plain-English AI glossary

In-distribution

● Core

Input examples that match the distribution your model was trained on.

Think of it like

Showing a dog-trained model another dog photo.

Example

If your model is trained on Wikipedia articles, another Wikipedia article is in-distribution. A tweet is not.

How it actually works

In-distribution examples get good model performance, but "distribution" is fuzzy. Is a dog photo taken at night in-distribution for a model trained on daytime photos? Probably partly. You can measure distributional similarity with Jensen-Shannon divergence or other metrics, but there's no sharp threshold. The practical question: does my model perform well on this type of input?

For product teams

Easier to predict model success. Challenges appear at distribution boundaries.

For engineers

Profile your training data distribution. Check new examples against it. Set an acceptance threshold.

Counterpart to out-of-distribution.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome