Neural Networks GMLE Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A data scientist is training a neural network to classify medical images. The training accuracy is 99%, but the validation accuracy is only 85%. The model is overfitting. Which of the following techniques would be most effective to reduce overfitting?

   • ARemove the validation set
   • BIncrease the number of hidden layers
   • CIncrease the learning rate
   • DAdd dropout to the hidden layers
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   A deep learning engineer is training a deep neural network with 20 hidden layers using sigmoid activation. The training loss decreases very slowly in the first few layers, while the later layers learn quickly. The engineer suspects vanishing gradients. Which of the following changes would most effectively address this issue while preserving the network's ability to model non-linear relationships?

   • AIncrease the learning rate
   • BReplace sigmoid with ReLU in all hidden layers
   • CUse batch normalization after each layer
   • DAdd L2 regularization
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   In a fully connected neural network layer, what is the role of the bias term?

   • AIt controls the learning rate during gradient descent.
   • BIt scales the input features before they are multiplied by the weights.
   • CIt prevents the activation function from saturating by clipping the neuron's output.
   • DIt allows the activation of a neuron to be shifted so the layer can represent functions that do not pass through the origin.
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   What is the purpose of early stopping in neural network training?

   • ATo reduce the learning rate when training stalls
   • BTo randomly drop neurons during training
   • CTo stop training when the validation loss starts to increase, preventing overfitting
   • DTo initialize weights to zero
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A neural network for sentiment analysis uses tanh activation in all hidden layers. The training loss decreases initially but then plateaus, and the gradients in the early layers are very small. The engineer wants to switch to an activation function that reduces the vanishing gradient problem while keeping the network's ability to model negative values. Which activation function should be used?

   • ASigmoid
   • BReLU
   • CLinear activation
   • DLeaky ReLU
   Reveal Solution Discussion Select an answer first