Model Drift
Model Drift
A phenomenon where an AI's predictions become less accurate because it can't keep up with real-world changes over time.
In Simple Terms
Model drift is what happens when an AI's prediction accuracy declines because the real-world situation has changed since it was trained. For example, when new slang or new shopping patterns emerge, an AI that only knows past data can no longer make accurate judgments. That's why systems like machine translation and fraud detection regularly check an AI's accuracy.
Behind the Name
The name comes from combining "model," referring to how an AI represents its predictions, with "drift," meaning to gradually move off course. It captures how an AI's predictions slowly drift away from reality as trends and lifestyles shift over time, leaving the model behind the world it's meant to represent.
Take a Closer Look!
Model drift is what happens when an AI that was built at one point in time gradually falls behind as the real world changes, causing its predictions and judgments to become less accurate.
Since AI learns patterns from past data, when people's lifestyles or the way society works shift, the AI can no longer come up with the right answers.
Simply put, there are two patterns behind model drift: changes in the input data and changes in the standard used to judge things.
For example, when a new trend causes the data being fed into an AI to change, that's called data drift.
On the other hand, when laws or social conditions change and the very definition of "what counts as correct" shifts, that's called concept drift.
If model drift is left unaddressed, an AI can keep making faulty predictions, which puts the reliability of the whole system at risk.
That's why in real-world operations, people continuously monitor an AI's predictions and periodically retrain it on fresh data to keep its accuracy up.