MLA-C01 Question 230
Select 2You are training a deep learning model on Amazon SageMaker, but the training loss does not decrease and the model fails to converge. Which of the following actions should you take to address the issue? (Choose TWO)
- A
Reduce the learning rate used by the optimizer.
- B
Increase the size of the training dataset.
- C
Switch the optimizer to a simpler one, such as stochastic gradient descent (SGD).
- D
Normalize or standardize the input features of the dataset.
- E
Increase the number of hidden layers in the neural network.
Show answer and explanation
Correct answers: A, D
Explanation
Convergence issues during training often arise due to improper learning rates or unscaled input features. Reducing the learning rate ensures that the optimizer takes smaller steps, preventing overshooting of the optimal solution. Normalizing or standardizing the input features ensures consistent scaling across features, helping the model converge more efficiently. Other options, such as increasing data size or adding layers, do not directly address convergence problems and may introduce additional complications.
- A. Correct.
Reducing the learning rate can help prevent the optimizer from overshooting the optimal solution, which is a common cause of convergence issues.
- B. Incorrect.
Increasing the size of the training dataset is not directly related to convergence issues. While more data can improve model generalization, it does not address problems during the optimization process.
- C. Incorrect.
Switching to a simpler optimizer like SGD is unlikely to resolve convergence problems, as advanced optimizers like Adam are better suited for handling complex optimization landscapes.
- D. Correct.
Normalizing or standardizing input features ensures that all features contribute equally to the optimization process, reducing the risk of poor convergence due to imbalanced feature scales.
- E. Incorrect.
Increasing the number of hidden layers may lead to overfitting or make the optimization process more challenging, potentially exacerbating convergence issues.