Google Professional Machine Learning Engineer Question 520
Select 3Google Cloud PlatformYou are tasked with deploying a machine learning model to production using Google Cloud AI Platform. After deployment, you notice that the model is returning consistent errors in predictions. Upon further inspection, you identify potential training and serving mismatches. Which of the following actions should you take to monitor and address this issue?
- A
Compare the feature distributions between the training dataset and the live serving dataset.
- B
Enable and monitor AI Platform Prediction request logs to analyze input data and identify anomalies.
- C
Retrain the model immediately using the same training dataset to eliminate potential serving issues.
- D
Set up alerting for data drift using Vertex AI Monitoring.
- E
Modify the model to include more layers and parameters for better prediction performance.
Show answer and explanation
Correct answers: A, B, D
Explanation
To address common training and serving errors, it's essential to monitor and compare the data distributions for any mismatches, enable logging to analyze prediction inputs, and use tools like Vertex AI Monitoring for proactive detection of data drift. These steps help ensure that the model performs as expected in production and that serving errors are minimized.
- A. Correct.
Correct. A mismatch between the training and serving data distributions is a common cause of serving errors. Comparing feature distributions can help identify inconsistencies.
- B. Correct.
Correct. Enabling prediction request logs allows you to monitor and analyze the inputs sent to the model during serving, which could help pinpoint anomalies causing errors.
- C. Incorrect.
Incorrect. Retraining the model with the same dataset without addressing the root cause of the issue is unlikely to resolve serving mismatches.
- D. Correct.
Correct. Vertex AI Monitoring can be used to detect data drift or feature skew in production, which are common causes of serving errors.
- E. Incorrect.
Incorrect. Modifying the model architecture without understanding the issue could lead to unnecessary complexity and may not address the root cause of the errors.