MLS-C01 Question 379
Select 2You are deploying a machine learning model using Amazon SageMaker, and the model's predictions are consistently inaccurate. You suspect that the issue could be related to the input features used during inference. Which of the following actions should you take to debug and resolve the issue? (Select TWO.)
- A
Use Amazon SageMaker Model Monitor to analyze the input features for drift during inference.
- B
Re-train the model with a larger training dataset to improve generalization.
- C
Verify that the features used during training and inference are preprocessed consistently.
- D
Use Amazon CloudWatch to check for errors in the model endpoint logs.
- E
Deploy the model on a GPU instance to increase prediction accuracy.
Show answer and explanation
Correct answers: A, C
Explanation
To debug inconsistent predictions, you must ensure that input features are processed consistently between training and inference, as mismatches can introduce errors. Amazon SageMaker Model Monitor is a powerful tool to detect data drift or inconsistencies in real-time input features, which is often a root cause of degraded model performance. Other options, like re-training with a larger dataset or deploying on a GPU instance, do not directly address the problem of feature inconsistencies.
- A. Correct.
Amazon SageMaker Model Monitor can detect drift in input features during inference by comparing real-time inputs to the baseline statistics generated during training, helping identify mismatches.
- B. Incorrect.
While re-training the model with a larger dataset can generally improve performance, it does not directly address issues related to inconsistent input features between training and inference.
- C. Correct.
Consistent preprocessing of input features during training and inference is critical. Mismatches in preprocessing can lead to inaccurate predictions.
- D. Incorrect.
Amazon CloudWatch can help debug endpoint issues like server errors but does not address feature-related inconsistencies directly.
- E. Incorrect.
Deploying on a GPU instance improves computation speed for inference but does not inherently improve model prediction accuracy.