A company deploys a computer vision model for quality inspection on a manufacturing line. After deployment, the model's accuracy drops from 95% to 80% over two weeks. Which action is most likely to address this issue?
Trap 1: Increase the confidence threshold for predictions.
This adjusts the trade-off between precision and recall but does not address the underlying data drift.
Trap 2: Decrease the learning rate of the training algorithm.
Learning rate is a hyperparameter for training, not for inference; it does not affect deployed model performance.
Trap 3: Deploy an additional ensemble of models for redundancy.
Ensemble methods improve accuracy if models are diverse, but they do not fix drift without retraining.
- A
Retrain the model using recently collected production data.
Retraining with current data adapts the model to new data distributions, countering drift.
- B
Increase the confidence threshold for predictions.
Why wrong: This adjusts the trade-off between precision and recall but does not address the underlying data drift.
- C
Decrease the learning rate of the training algorithm.
Why wrong: Learning rate is a hyperparameter for training, not for inference; it does not affect deployed model performance.
- D
Deploy an additional ensemble of models for redundancy.
Why wrong: Ensemble methods improve accuracy if models are diverse, but they do not fix drift without retraining.