Google Professional Machine Learning Engineer Question 86
Single answerGoogle Cloud PlatformYou have trained a machine learning model on Google Cloud AI Platform, and the model is deployed to an endpoint for serving. During testing, you notice that the predictions are consistently incorrect for a specific input feature range. What is the most appropriate next step to debug and resolve the issue?
- A
Analyze the input data for potential mismatches between training and serving feature distributions.
- B
Retrain the model using a larger batch size to improve generalization.
- C
Increase the number of epochs during training to ensure the model learns the feature patterns.
- D
Enable Explainable AI in Google Cloud AI Platform to understand feature importance and identify potential issues.
Show answer and explanation
Correct answer: A
Explanation
The most effective way to debug and resolve the issue of incorrect predictions is to first analyze the input data for discrepancies between the training and serving feature distributions. This step ensures that the model is being served with data that matches the conditions it was trained on. If discrepancies are found, appropriate data preprocessing, retraining, or feature engineering steps can be taken to resolve the issue.
- A. Correct.
Analyzing the input data for mismatches is the correct approach because feature distribution discrepancies between training and serving are a common cause of poor model performance. Identifying these mismatches can help rectify the issue, such as by reprocessing the data or retraining the model.
- B. Incorrect.
Retraining the model with a larger batch size does not specifically address the issue of incorrect predictions for a specific feature range. Batch size primarily affects training stability and convergence, not feature-specific issues.
- C. Incorrect.
Increasing the number of epochs may lead to overfitting and does not directly address the underlying issue of incorrect predictions caused by feature distribution problems.
- D. Incorrect.
While enabling Explainable AI can provide insights into feature importance, it is not the most immediate or targeted action for resolving the issue of incorrect predictions due to feature discrepancies.