Google Professional Machine Learning Engineer Question 501
Select 3Google Cloud PlatformYou are working on a machine learning pipeline deployed on Google Cloud, and you want to monitor for training-serving skew between your training data and the data used in production. Which strategies should you implement to detect and mitigate this issue effectively?
- A
Use Dataflow to compute statistics on production data and compare them with training data statistics.
- B
Implement Vertex AI Feature Store to track and monitor feature distributions over time.
- C
Create a model evaluation pipeline to compare the accuracy of the training model and the production model predictions.
- D
Log predictions and input features in production, and compare them with the training dataset features using TensorFlow Data Validation (TFDV).
- E
Enable the AI Explanations feature in Vertex AI to analyze prediction differences between training and serving data.
Show answer and explanation
Correct answers: A, B, D
Explanation
Training-serving skew occurs when the data distributions in training and production differ, which can lead to degraded model performance. To monitor for this issue, you can leverage tools like Dataflow to compute statistics on production data, use Vertex AI Feature Store to track feature distributions over time, and log production data to compare with training data using TensorFlow Data Validation (TFDV). These strategies help identify and mitigate skew effectively, ensuring the model performs reliably in production.
- A. Correct.
Correct: Dataflow can be used to process and compute statistics on production data in real-time, enabling comparisons with training data statistics to detect skew.
- B. Correct.
Correct: Vertex AI Feature Store provides a centralized repository to manage and monitor feature distributions over time, helping identify divergence between training and serving data.
- C. Incorrect.
Incorrect: While model evaluation pipelines are useful for monitoring model performance, they do not directly help in detecting training-serving skew, which focuses on data mismatches rather than accuracy differences.
- D. Correct.
Correct: Logging predictions and input features from production, and comparing them with training data using tools like TFDV, is a common and effective way to detect training-serving skew.
- E. Incorrect.
Incorrect: AI Explanations in Vertex AI is used to interpret model predictions, but it does not directly address the detection of training-serving skew.