4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

Quiz Test (10 Marks): Deep Learning

Note: Attempt all questions in the quiz.

1. Email *

2. A 2D input of size 128×128 is passed through 3 convolutional layers with 1 point

kernel size 3, stride 2, and padding 1. What is the final output size?

3. In a Transformer with 8 heads, model dimension 512, what is the 1 point

dimensionality of each head?

4. If your CNN model takes input 224x224x3 and passes it through a 7x7 conv 1 point

with stride 2, padding 3, and 64 filters, what is the output shape?

5. A GRU cell has input size 20 and hidden size 10. How many parameters 1 point

does it have?

6. A model predicts [0.8, 0.1, 0.1] for class labels, but the true label is [1, 0, 0]. 1 point

What is the categorical cross-entropy loss?

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 1/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

7. An LSTM has input size 100 and hidden size 50. How many parameters are 1 point
in the LSTM cell?

8. A ReLU activation function outputs 0 for how many inputs out of these: [-5, 1 point
0, 3, -2, 7]?

9. A model has 2 hidden layers with 128 and 64 neurons respectively. Each 1 point
layer has a bias and ReLU activation. If the input vector is of size 256,
calculate the total number of parameters (excluding output layer).

10. Dropout is used in neural networks to: 1 point

Mark only one oval.

Speed up inference

Reduce memory usage

Avoid vanishing gradients

Prevent overfitting

11. In an LSTM cell, what does the forget gate do? 1 point

Mark only one oval.

Selectively updates the cell state

Outputs the next hidden state

Deletes all previous memory

Determines what information to discard

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 2/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

12. Which optimizer adapts learning rates based on the first and second 1 point

moments of the gradients?

Mark only one oval.

SGD

Adagrad

Adam

Momentum

13. What is the primary benefit of using batch normalization in deep 1 point
networks?

Mark only one oval.

Prevent overfitting

Speed up training and stabilize learning

Reduce the number of layers

Increase model complexity

14. Transformers eliminate which of the following traditionally used 1 point

Mark only one oval.

Convolutions

Attention mechanisms

Recurrent layers

Embeddings

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 3/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

15. Layer normalization is preferred over batch normalization in: 1 point

Mark only one oval.

CNNs with large batches

Transformers and NLP tasks

GANs

Reinforcement Learning only

16. The purpose of Xavier initialization is to: 1 point

Mark only one oval.

Reduce dropout variance

Prevent weight explosion/vanishing in deep networks

Make loss function convex

Regularize the output layer

17. A model’s training loss keeps decreasing, but validation loss starts 1 point
increasing after epoch 15. What does this indicate and what should you
do?

18. You are given a batch of input values: [10,12,14,16]. 1 point

The Batch Normalization layer computes the normalized values using the
batch mean and standard deviation. Then, it applies learnable parameters
γ=2 and β=5 as follows: What is the output of the Batch Normalization
layer after applying γ and β?

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 4/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

19. A deep learning model is used for semantic segmentation on medical 1 point

imaging data. The ground truth and prediction for a particular image are
binary masks (1 for foreground, 0 for background).

You are given:

Ground Truth Mask (GT):

Predicted Mask (Pred):

What is the Intersection over Union (IoU) for the foreground class?

20. A deep learning model is used for semantic segmentation on medical 1 point
imaging data. The ground truth and prediction for a particular image are
binary masks (1 for foreground, 0 for background).

You are given:

Ground Truth mask:

Predicted mask:
[1,0,0,1,0,1,1,0]

What is the Dice Coefficient for the foreground class (1)?

21. Suppose your model gives the raw logits for 3 classes at a pixel: 1 point

Logits=[2.0,1.0,0.1]

What is the SoftMax probability for each class (to 2 decimal places)?

This content is neither created nor endorsed by Google.

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 5/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

Forms

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 6/7
4/9/25, 12:59 PM Quiz Test (10 Marks): Deep Learning and Applications CSBB 424

https://docs.google.com/forms/d/1JlAOVX2js6Eq8iGeuFsP8shFQ7fhMElASeNewQ_vLik/edit 7/7