This video breaks down the differences between data drift and model drift.
One of the elements of an MLOps pipeline is drift detection, which helps you monitor model performance over time and identify when it's time for retraining. Once you've successfully deployed models and are running live inferences in production, you'll encounter yet another obstacle: monitoring model performance over time. We monitor several model performance indicators, including overall scoring, inference speed, latency, accuracy, and finally data drift and model drift. This tech talk covers the algorithms we've developed to automate detection of data drift and model drift, or input and output drift.