AI Aging

Found this article by NannyML and thought it would be useful for my article writing.

ML models depend on the data it was trained on. Since data in the real-world changes over time, model performance will degrade as time passes, a phenomenon the authors called "AI Aging".

MIT, Harvard, and other top institutions trained 4 different ML models (Linear Regression, Random Forest Regressor, XGBoost, and a Multilayer Perceptron Neural Network) on 32 datasets from four industries (Healthcare, Weather, Airport Traffic, and Financial) and found that 91% of their ML models degrade over time

Key takeaway

Neither the data nor the model alone can be used to guarantee consistent predictive quality. Instead, the temporal model quality is determined by the stability of a specific model applied to the specific data at a particular time.

What are the solutions?

The right solution is context-dependent, and investigation should be done to understand the cause of the degradation

The solutions are:

  1. Alert when model must be retrained: need access to latest ground truth or able to estimate model performance
  2. develop efficient and robust mechanism for automatic model retraining: (if no data quality issue or concept drift), retraining model on latest labeled data can help
  3. Have constant access to most recent ground truth: allows retraining, but in practice, ground truth is often delayed, expensive and time-consuming to newly labeled data. Alternative is to have a model catalog and use estimated performance to select best performing model

More:

  • https://huyenchip.com/2022/02/07/data-distribution-shifts-and-monitoring.html
  • https://www.nannyml.com/blog/6-ways-to-address-data-distribution-shift

4/17/2023