Glossary · AI engineering and governance
Training-serving skew
Also known as: Train-serve skew
German: Abweichung zwischen Training und Betrieb
In machine learning operations, training-serving skew is a difference between the data or processing a model sees during training and what it receives in operation, for example because features are computed differently, sensors are scaled differently or preprocessing code differs, leading to worse performance than validated.
- Industrial AI
- AI
In one sentence
Training-serving skew is a mismatch between training data or processing and what a model receives in operation, degrading performance.
Example
In training, the temperature feature was averaged over 10 seconds; on the edge device it is sampled instantly, so the deployed model sees noisier inputs than expected.
How it applies
- Engineering: Use the same code for feature computation in training and operation wherever possible, and log the inputs the deployed model actually receives.
- Commissioning: Before activating a model, compare a sample of live inputs with the training data: units, scaling, sampling rates, time alignment and missing-value handling.
- Documentation: Specify the model's input interface precisely: signals, units, sampling, preprocessing and valid ranges. This specification is what integrators need to avoid skew.
Training-serving skew vs. data drift
Training-serving skew exists from the moment of deployment because of a technical mismatch. Data drift develops over time because the real process changes. Skew is fixed by correcting the pipeline; drift may require retraining. A short comparison of logged live inputs against the training data during commissioning catches most causes of skew.