Back to Study Material Machine Learning

Retraining a model

Retraining a model means updating its parameters using new data to restore or improve performance that has degraded due to drift or to incorporate newly av...

Retraining a model

Updating a Model as the World Changes

Retraining a model means updating its parameters using new data to restore or improve performance that has degraded due to drift or to incorporate newly available information. A model is not a one-time artifact: it is a living component that must be refreshed as the world it models evolves.

Retraining strategies vary along a spectrum. Full retraining discards the old model entirely and trains from scratch on all available data, including historical and recent examples. This is the most thorough approach but also the most expensive. Incremental learning updates the existing model with only the new data, adjusting weights without processing historical examples again. It is faster but risks catastrophic forgetting, where learning new patterns overwrites previously learned ones.

Fine-tuning is common with neural networks and large pretrained models. The model's weights are used as a starting point and only the later layers or all layers are updated with new data using a small learning rate. This preserves general representations learned from large historical datasets while adapting to recent patterns.

Retraining triggers can be time-based, such as weekly or monthly updates, or event-based, triggered by monitoring alerts when performance metrics cross a threshold. Automated pipelines that collect new labelled data, validate quality, retrain, evaluate on a holdout set, and promote the new model to production only if it outperforms the current one are the industry standard for high-stakes applications.

A concrete example: a loan default prediction model was retrained quarterly. After a sudden interest rate increase, monitoring showed accuracy dropping from 88 to 79 percent within two weeks. An event-based trigger fired a retraining pipeline using the three months of post-increase data, and the model recovered to 86 percent accuracy.

Exam tip: know that retraining is not always about adding more data — sometimes retraining with a data window that excludes very old examples improves performance because recent patterns are more relevant than historical ones.