MLA-C01 Practice Question: A data scientist is using Amazon SageMaker…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Define a Debugger rule that monitors the loss plateau

  Why it's wrong here

  A loss plateau rule detects stalled loss, which signals underfitting or slow convergence, not overfitting. It is tempting because it is a genuine Debugger rule that can stop training, but overfitting is caught by rules comparing training and validation loss, such as overfit.

• ✓

  Configure a custom rule that triggers a STOP training action when validation loss stops decreasing

  Why this is correct

  A custom Debugger rule evaluates the validation-loss tensor against a threshold and can emit a StopTraining action, halting the job automatically. This satisfies the requirement to stop training without manual intervention when validation loss ceases to decrease, indicating overfitting.

• ✗

  Create a SageMaker Training Compiler

  Why it's wrong here

  Training Compiler reduces GPU memory use and accelerates training; it takes no Debugger rule input and cannot halt a job on overfitting. It is tempting because it optimises training runs, but stopping on a metric condition requires a Debugger rule with a stop action.

• ✗

  Use a built-in rule that checks for vanishing gradients

  Why it's wrong here

  Vanishing gradients indicate unstable backpropagation, not overfitting, so this rule would never fire on the stated symptom. It is tempting because SageMaker Debugger's built-in rules do cover gradient problems, and such a rule would be the right pick if training stalled or diverged rather than overfit.

Go deeper

Related to this question