Glossary · AI engineering and governance
Model validation
German: Modellvalidierung
In machine learning, model validation is the evaluation of a trained model against defined acceptance criteria with data not used for training, to confirm that it is fit for its intended purpose under the intended operating conditions. It includes performance metrics, robustness checks and review of known limitations.
- Industrial AI
- AI
In one sentence
Model validation checks a trained model against acceptance criteria with unseen data to confirm it fits its intended use.
Example
Before release, a weld inspection model is tested on 2,000 images from a third line it has never seen and must detect at least the agreed share of defects.
How it applies
- Engineering: Define acceptance criteria before testing: metrics, thresholds, test data and relevant operating conditions. Test data should represent real use, including rare but important cases.
- Change management: Revalidate after retraining, changes to input data sources, or changes to the process. A retrained model is a new version (Model versioning).
- Documentation: A validation report records the model version, test data, criteria, results, limitations and approval. This is part of the evidence for the AI system and feeds the user information on its limits.
Model validation vs. validation in systems engineering
In machine learning, validation often also names a data split used to tune hyperparameters, which is different from final testing. In systems engineering, Validation confirms that the system meets the user's needs in its intended use, while Verification confirms that it meets its specification. A validated model still needs validation of the overall system it is part of.