uiz Space

May 2026 term · Introduction to Deep Learning and Generative AI · BSDA2001

Introduction to Deep Learning and Generative AI Quiz 2: 16 August 2026 (May 2026 term)

The IIT Madras BS Introduction to Deep Learning and Generative AI (Deep Learning and GenAI) Quiz 2 paper sat on 16 Aug 2026, in the May 2026 term: 19 questions for 47 marks in 120 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.

Questions
19
Marks
47
Duration
120 min
MCQ
10
MSQ
6
Numerical
3
Take as mock testPractise with answers

Updated

Official paper: Introduction To Deep Learning And Generative Ai 14 Aug 26 · No negative marking.

Question 1

+2 marksOne correct option

Consider the following code snippet:

Based on the standard normalization technique for a single-channel grayscale image dataset like FashionMNIST, which of the following correctly fills in blank (A)?

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • A
Question 1 on its own page

Question 2

+2 marksOne correct option

Suppose during training Epoch 1 D(real)=0.90 D(fake)=0.10 Later Epoch 20 D(real)=0.999 D(fake)=0.0001
What problem is most likely? (Here, D(x) represents Discriminator's output on image x)

  1. A

    Discriminator has become too strong, leading to vanishing gradients for the generator.

  2. B

    Mode collapse has definitely occurred.

  3. C

    Generator has perfectly learned the data distribution.

  4. D

    The discriminator will stop updating permanently.

Show answer

Correct answer

  • A

    Discriminator has become too strong, leading to vanishing gradients for the generator.

Question 2 on its own page

Question 3

+2 marksOne correct option

Let D(x) denote the discriminator's output. Which of the following represents the most desirable discriminator output during stable GAN training?

  1. A

    D(real) = 0.9, D(fake) = 0.1

  2. B

    D(real) = 0.8, D(fake) = 0.2

  3. C

    D(real) = 1.0, D(fake) = 0.0

  4. D

    D(real) = 0.5, D(fake) = 0.5

Show answer

Correct answer

  • D

    D(real) = 0.5, D(fake) = 0.5

Question 3 on its own page

Question 4

+2 marksOne correct option

What is the output of the following code snippet?

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • D
Question 4 on its own page

Question 5

+2 marksOne correct option
  1. A
  2. B
  3. C

    Truncation eliminates the vanishing-gradient problem entirely, because gradients no longer traverse long products.

  4. D

    Truncation increases memory usage because the whole computation graph must be retained.

Show answer

Correct answer

  • B
Question 5 on its own page

Question 6

+2 marksOne correct option

Consider the following implementation of a RNN:

What is the output of the print statements?

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • A
Question 6 on its own page

Question 7

+2 marksOne correct option

Which of the following options is correct regarding the standard encoder-decoder architecture?

  1. A

    During training, we feed the decoder's prediction from the previous step as input for the current step to stabilize learning.

  2. B

    We first train the encoder completely, and then using the context vector from the fully trained encoder, we train the decoder.

  3. C

    When working with batches of varying lengths, the loss function calculates gradients for padding tokens to ensure uniform batch updates.

  4. D

    None of these

Show answer

Correct answer

  • D

    None of these

Question 7 on its own page

Question 8

+3 marksOne correct option
  1. A
  2. B

    (A figure from the original paper is missing from the source site.)

  3. C
  4. D
Show answer

Correct answer

  • B

    (A figure from the original paper is missing from the source site.)

Question 8 on its own page

Question 9

+3 marksOne correct option

Which of the following statements must be true?

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • C
Question 9 on its own page

Question 10

+3 marksOne correct option

A GAN is trained on images preprocessed using the following PyTorch transformation:

Which of the following activation functions are most suitable for the final layers of the Generator and Discriminator, respectively?

  1. A

    Tanh, Sigmoid

  2. B

    Sigmoid, Tanh

  3. C

    ReLU, Sigmoid

  4. D

    Tanh, Tanh

  5. E

    Sigmoid, Sigmoid

  6. F

    ReLU, ReLU

Show answer

Correct answer

  • A

    Tanh, Sigmoid

Question 10 on its own page

Question 11

+2 marksOne or more correct options

Select the correct statements from the following:

Select all that apply.

  1. A

    If the sequence length doubles while all model hyperparameters remain unchanged, the number of trainable parameters in the RNN also doubles.

  2. B
  3. C

    The same input-to-hidden and hidden-to-hidden weight matrices are reused at every time step, regardless of the sequence length.

  4. D

    An RNN can process sequences of arbitrary length only if the input size changes with the sequence length.

Show answer

Correct answers

  • B
  • C

    The same input-to-hidden and hidden-to-hidden weight matrices are reused at every time step, regardless of the sequence length.

Question 11 on its own page

Question 12

+2 marksOne or more correct options

Which of the following increase(s) the output size?

Select all that apply.

  1. A

    Increasing kernel size

  2. B

    Increasing padding

  3. C

    Increasing stride

Show answer

Correct answer

  • B

    Increasing padding

Question 12 on its own page

Question 13

+2 marksOne or more correct options

Select all statements that are TRUE.

Select all that apply.

  1. A
  2. B

    The intersection of two convex sets is convex, but the union of two convex sets need not be.

  3. C

    If every local minimum of a function is also a global minimum, the function must be convex.

  4. D
  5. E
Show answer

Correct answers

  • A
  • B

    The intersection of two convex sets is convex, but the union of two convex sets need not be.

  • D
  • E
Question 13 on its own page

Question 14

+4 marksOne or more correct options

Select all that apply.

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answers

  • A
  • C
Question 14 on its own page

Question 15

+3 marksOne or more correct options

Select all that apply.

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answers

  • B
  • C
  • D
Question 15 on its own page

Question 16

+3 marksOne or more correct options

Select all statements that are TRUE regarding the mechanics of Adagrad and RMSProp.

Select all that apply.

  1. A

    Assuming a constant base learning rate, Adagrad's squared-gradient accumulator only grows (or stays the same), meaning a parameter's effective learning rate can never increase over time.

  2. B

    In RMSProp, a parameter's effective learning rate can actually increase again if it starts receiving much smaller gradients than it did previously.

  3. C

    For sparse features (which rarely receive non-zero gradients), Adagrad forces the effective learning rate to decay much more slowly compared to frequently updated dense features.

  4. D

    Over a very long training run, RMSProp's accumulator eventually converges to the total sum of all squared gradients seen during training.

  5. E

    Changing Adagrad's infinite cumulative sum into an exponentially decaying average is exactly the modification that creates RMSProp and solves the vanishing learning rate problem.

Show answer

Correct answers

  • A

    Assuming a constant base learning rate, Adagrad's squared-gradient accumulator only grows (or stays the same), meaning a parameter's effective learning rate can never increase over time.

  • B

    In RMSProp, a parameter's effective learning rate can actually increase again if it starts receiving much smaller gradients than it did previously.

  • C

    For sparse features (which rarely receive non-zero gradients), Adagrad forces the effective learning rate to decay much more slowly compared to frequently updated dense features.

  • E

    Changing Adagrad's infinite cumulative sum into an exponentially decaying average is exactly the modification that creates RMSProp and solves the vanishing learning rate problem.

Question 16 on its own page

Question 17

+3 marksNumerical answer

Calculate the total number of parameters in the RNN implemented in the following code:

Show answer

Correct answer: 5504

Question 17 on its own page

Question 18

+3 marksNumerical answer
Show answer

Correct answer: 117248

Question 18 on its own page

Question 19

+2 marksNumerical answer
Show answer

Correct answer: 100

Question 19 on its own page