DataTerminology.comNEW
Also known as: Model Refresh
Updating a machine learning model with new data to maintain performance as distributions, requirements, or contexts change over time.
DESCRIPTION
Retraining refreshes models when data drifts or objectives evolve. It can be periodic, triggered by monitoring thresholds, or continuous via pipelines. Effective retraining includes data validation, feature stability checks, and reproducible experiments. Governance covers versioning, approvals, and rollback plans. A/B testing and shadow deployments evaluate changes safely. Automation reduces toil and speeds adaptation.
ALTERNATIVE DEFINITIONS
Microsoft Learn
Operationalization includes monitoring models and retraining them when performance degrades.
Places retraining within monitored operational ML lifecycle.
Neptune.ai
Model retraining is the process of updating a machine learning model with new data to sustain its performance.
Defines retraining with a focus on performance maintenance.
Google Cloud
Continuous delivery pipelines automate data and model training, evaluation, and deployment, enabling regular retraining.
Frames retraining as part of automated ML pipelines.
RELATED TERMS
Model Drift Data Drift Continuous Training Model Monitoring Concept Drift Workflow Orchestration Model Versioning Feature Store A/B Testing
COMPETING TERMS
KNOWLEDGE AREAS
MLOps AI Data Science Observability & Telemetry
TARGET AUDIENCE
MLOps Engineer Data Scientist AI Engineer Chief AI Officer
EXTERNAL SOURCES
- Wikipedia - Concept drift
- Google Cloud - MLOps pipelines for continuous retraining
- Microsoft Learn - Model monitoring triggers retraining
- AWS - Monitoring drift for retraining
TOP QUESTIONS
When should Retraining be triggered?
Trigger when performance metrics degrade, drift is detected, or business requirements change.
Category: Management
What are common mistakes in Retraining?
Skipping data validation, failing to compare against baselines, and deploying without shadow testing.
Category: Mistakes
How does Retraining relate to concept drift?
Retraining combats drift by adapting models to new data patterns and distributions.
Category: Definition
What is the risk of frequent Retraining?
Overfitting to recent noise, increased costs, and instability; use robust triggers and validation.
Category: Risks
How to validate a retrained model?
Use holdout tests, cross-validation, and A/B or canary deployments to compare performance and risks.
Category: Testing
What infrastructure supports automated Retraining?
Pipelines, feature stores, registries, and observability tools integrated with CI/CD and orchestration.
Category: Requirements
Is fine-tuning different from Retraining?
Fine-tuning adjusts model parameters on new data; retraining may rebuild from scratch or update incrementally.
Category: Misunderstanding
How to govern Retraining?
Version models, record experiment metadata, approve deployments, and maintain rollback plans.
Category: Management
What metrics indicate successful Retraining?
Improved accuracy, calibration, fairness metrics, and stable latency/cost under production traffic.
Category: Testing