Google Professional Machine Learning Engineer Question 518
Select 3Google Cloud PlatformYou are a machine learning engineer responsible for deploying a model to production on Google Cloud. After deployment, you notice that the model's predictions are increasingly inaccurate over time. Upon investigating, you identify a mismatch between the data distribution during training and the data being served in production. Which actions should you take to monitor and address such issues effectively?
- A
Enable Vertex AI Model Monitoring to track feature skew and data drift in production.
- B
Retrain the model immediately with the new production data to fix the issue.
- C
Set up alerting for anomalies in production data distribution compared to training data.
- D
Implement logging to capture input features and predictions for further analysis.
- E
Modify the input data pipeline to enforce the same distribution as the training data.
Show answer and explanation
Correct answers: A, C, D
Explanation
Monitoring for training-serving discrepancies is critical for maintaining machine learning model performance in production. Using Vertex AI Model Monitoring, anomaly alerting, and logging enables proactive detection and diagnosis of issues like feature skew and data drift. These measures allow engineers to make informed decisions about retraining or other corrective actions while ensuring the model remains reliable and accurate in production.
- A. Correct.
Correct: Vertex AI Model Monitoring is designed to track feature skew, data drift, and other issues in production, helping detect mismatches between training and serving data distributions.
- B. Incorrect.
Incorrect: While retraining the model may eventually be necessary, it is not the immediate action to monitor or address the issue effectively. Proper monitoring and diagnostics should precede retraining.
- C. Correct.
Correct: Alerting for anomalies helps in early detection of data drift and feature skew, allowing for timely intervention before the model's performance significantly degrades.
- D. Correct.
Correct: Logging input features and predictions is crucial for diagnosing the root cause of the issue and identifying specific features contributing to the drift.
- E. Incorrect.
Incorrect: Modifying the input data pipeline to enforce the same distribution as the training data is impractical and can lead to inaccurate results since production data often evolves naturally over time.