A deep learning engineer is training a very deep network with batch normalization. The training loss decreases rapidly at first but then becomes unstable, with occasional spikes. The engineer is using a small batch size of 8. What is the most likely cause of this instability?

The activation function is saturating.

The small batch size makes the batch statistics noisy, leading to unstable normalization.

What is the primary role of backpropagation in training a neural network?

To compute the gradient of the loss with respect to each weight

To initialize the weights of the network randomly

To forward propagate inputs through the network to produce outputs

To reduce the number of neurons in the network during training

A neural network with two hidden layers is trained using backpropagation. The loss decreases for a while, then the gradients in the first hidden layer become extremely small, and the weights stop updating effectively. The network uses sigmoid activation functions. Which change would most directly address this issue?

Use mean squared error instead of cross-entropy.

Replace sigmoid activations with ReLU in the hidden layers.

What is the primary benefit of batch normalization in deep neural networks?

It eliminates the need for activation functions in hidden layers

It reduces the number of parameters in the network, making it lighter

It stabilizes and accelerates training by normalizing layer inputs

It ensures the model never overfits regardless of the dataset size

A deep learning practitioner is training a very deep network. The gradients are becoming extremely small in the early layers, making training ineffective. The practitioner wants to use a technique that normalizes the inputs to each layer to help mitigate this issue. Which technique should be used?

Weight initialization with zeros

CompTIADataX

Domain 3Objective 4

Deep Learning DY0-001 Practice Questions (Page 2)

Part of the Machine learning domain, which accounts for 24% of the DY0-001 exam. CompTIA does not publish an official question count, but from its 165-minute exam (~65–110 total, ~16–26 in this domain), expect 3–5 from this objective — we provide 24 practice questions to prepare you well beyond it. (estimate)

24questions here

5free pages

5concepts

24%of the exam

Questions 6–10

6expert · hard
A deep learning engineer is training a very deep network with batch normalization. The training loss decreases rapidly at first but then becomes unstable, with occasional spikes. The engineer is using a small batch size of 8. What is the most likely cause of this instability?
Select an answer first
7foundation · easy
What is the primary role of backpropagation in training a neural network?
Select an answer first
8expert · hard
A neural network with two hidden layers is trained using backpropagation. The loss decreases for a while, then the gradients in the first hidden layer become extremely small, and the weights stop updating effectively. The network uses sigmoid activation functions. Which change would most directly address this issue?
Select an answer first
9foundation · easy
What is the primary benefit of batch normalization in deep neural networks?
Select an answer first
10application · medium
A deep learning practitioner is training a very deep network. The gradients are becoming extremely small in the early layers, making training ineffective. The practitioner wants to use a technique that normalizes the inputs to each layer to help mitigate this issue. Which technique should be used?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to DY0-001

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by CompTIA. “DY0-001” is a trademark of its owner, used for identification only.