Neural Networks GMLE Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   Which loss function is most appropriate for a binary classification task where the model outputs a probability?

   • AMean Squared Error (MSE)
   • BHinge Loss
   • CCategorical Cross-Entropy
   • DBinary Cross-Entropy
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A machine learning engineer is tuning a neural network for a regression task. The model currently uses a batch size of 32 and a learning rate of 0.01 with SGD. The training loss is decreasing but the validation loss is noisy. The engineer wants to improve the stability of validation loss without significantly increasing training time. Which of the following changes should be made?

   • AIncrease the batch size to 128
   • BDecrease the batch size to 8
   • CSwitch to a shallower network
   • DIncrease the learning rate to 0.1
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A machine learning engineer is training a neural network on a large dataset. The loss curve is noisy and the model takes a long time to converge. The engineer wants to speed up training while maintaining stability. Which of the following combinations of changes would be most effective?

   • AIncrease learning rate and use SGD
   • BIncrease batch size and use Adam optimizer
   • CDecrease learning rate and use Adam
   • DDecrease batch size and use SGD
   Reveal Solution Discussion Select an answer first
5. 19expert · medium

   A machine learning engineer is training a neural network with a large dataset. The loss is decreasing but the training is slow. The engineer wants to use an optimizer that adapts the learning rate for each parameter and also incorporates momentum. Which optimizer should be used?

   • AAdam
   • BSGD
   • CRMSProp
   • DAdagrad
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A deep learning engineer is training a neural network with a hidden layer using tanh activation. The gradients are computed via backpropagation. The engineer notices that for some inputs, the gradients are very close to zero, causing the weights to stop updating. Which of the following is the most likely cause and the best fix?

   • AThe network is overfitting; add dropout
   • BThe learning rate is too high; reduce it
   • CThe loss function is incorrect; use a different loss
   • DThe tanh activation saturates for large inputs; switch to ReLU
   Reveal Solution Discussion Select an answer first