MLA-C01 Question 231
Select 3You are training a deep learning model on a large dataset using Amazon SageMaker. After several epochs, the model's loss remains stagnant, and the validation accuracy does not improve. You suspect a convergence issue. Which of the following actions should you take to address this problem?
- A
Reduce the learning rate to allow the model to converge more gradually.
- B
Increase the size of the training batch to improve gradient estimation.
- C
Normalize or scale the input features to ensure numerical stability.
- D
Switch to a simpler model architecture to reduce overfitting.
- E
Check for potential data quality issues, such as mislabeled samples or outliers.
Show answer and explanation
Correct answers: A, C, E
Explanation
Convergence issues can arise due to several factors, including inappropriate learning rates, unscaled input data, or poor-quality training data. Reducing the learning rate can help the optimizer find the minima more effectively, while normalizing the input data ensures numerical stability during training. Additionally, verifying data quality is crucial to rule out problems that could hinder learning.
- A. Correct.
Reducing the learning rate can help the optimizer take smaller steps toward convergence, which is beneficial when the loss is not decreasing.
- B. Incorrect.
Increasing the batch size does not directly address convergence issues. In some cases, larger batch sizes may even hinder convergence by reducing gradient variability.
- C. Correct.
Normalizing or scaling input features ensures that the model receives numerically stable data, which can aid in convergence.
- D. Incorrect.
Switching to a simpler model architecture is more relevant for addressing overfitting, not convergence issues.
- E. Correct.
Data quality issues, such as mislabeled samples or outliers, can prevent the model from converging effectively and should be investigated.