uiz Space

September 2025 term · Introduction to Deep Learning and Generative AI · BSDA2001

Introduction to Deep Learning and Generative AI End Term: 21 December 2025 (September 2025 term)

The IIT Madras BS Introduction to Deep Learning and Generative AI (Deep Learning and GenAI) End Term paper sat on 21 Dec 2025, in the September 2025 term: 30 questions for 97 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.

Questions
30
Marks
97
Duration
180 min
MCQ
21
Numerical
7
MSQ
2
Take as mock testPractise with answers

Updated

Official paper: Introduction To Deep Learning And Generative Ai 18 Dec 25 · No negative marking.

Question 1

+3 marksOne correct option

You are training an artificial neural network (ANN) for a regression task to predict house prices. After inspecting the target variable (price) in your dataset, you observe that while most prices fall within a predictable range, there are a handful of extremely expensive mansions and a few properties sold under unusual circumstances that represent severe outliers. Your primary goal is to build a model that performs well on typical houses without letting its predictions be overly influenced by these rare, extreme values. You also desire smoother convergence during training. Which loss function should you primarily consider for this task?

  1. A

    Mean Squared Error

  2. B

    Sum of Squared Errors

  3. C

    Huber Loss

  4. D

    Binary Cross-Entropy Loss

Show answer

Correct answer

  • C

    Huber Loss

Question 1 on its own page

Question 2

+5 marksNumerical answer
Show answer

Correct answer: 0.5 (accepted within ±0.1)

Question 2 on its own page

Question 3

+3 marksOne or more correct options

Which of the following are true about encoding categories?

Select all that apply.

  1. A

    One-hot encoding represents each class as a dense vector that captures semantic relationships.

  2. B

    Embedding layers always assign a fixed manually-defined vector to each class.

  3. C

    For any two distinct categories represented using one-hot encoding, the dot product between their vectors is always 0 .

  4. D

    For any two distinct categories represented using one-hot encoding, the Euclidean distance between their vectors is always 1 .

  5. E

    Embedding layers learn meaningful dense vectors where similar concepts have similar representations.

Show answer

Correct answers

  • C

    For any two distinct categories represented using one-hot encoding, the dot product between their vectors is always 0 .

  • E

    Embedding layers learn meaningful dense vectors where similar concepts have similar representations.

Question 3 on its own page

Question 4

+2 marksNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 6 (accepted within ±0.1)

Question 4 on its own page

Question 5

+5 marksNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 2 (accepted within ±0.1)

Question 5 on its own page

Question 6

+4 marksOne correct option
  1. A

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

  2. B
  3. C
  4. D
Show answer

Correct answer

  • B
Question 6 on its own page

Question 7

+3 marksOne or more correct options

Select all that apply.

  1. A
  2. B
  3. C
  4. D
  5. E
  6. F
Show answer

Correct answers

  • A
  • C
  • E
Question 7 on its own page

Question 8

+2 marksOne correct option
  1. A

    There is no error; this correctly replaces the final layer.

  2. B

    The entire model.classifier Sequential module must be replaced, not just a single layer.

  3. C

    The new nn.Linear layer's in_features is incorrect; it should be 1280 to match the features coming from the rest of the network, not 1000.

  4. D

    The dropout layer at model.classifier[0] must be removed for transfer learning to work.

Show answer

Correct answer

  • C

    The new nn.Linear layer's in_features is incorrect; it should be 1280 to match the features coming from the rest of the network, not 1000.

Question 8 on its own page

Question 9

+2 marksOne correct option
  1. A

    The Normalize transform should not use ImageNet statistics for food images.

  2. B

    The Normalize transform is applied before ToTensor, but it expects a tensor with values in the [0, 1] range as input.

  3. C

    The ToTensor transform should be the first operation, applied directly to the file path, not the PIL Image object.

  4. D

    There is no issue; the pipeline correctly converts and normalizes the image.

Show answer

Correct answer

  • B

    The Normalize transform is applied before ToTensor, but it expects a tensor with values in the [0, 1] range as input.

Question 9 on its own page

Question 10

+3 marksNumerical answer

A single Inception module has four parallel branches. After their respective convolutions, the branches produce feature maps with 64, 128, 32, and 32 channels. What is the total number of channels in the module's output feature map?

Show answer

Correct answer: 256

Question 10 on its own page

Question 11

+4 marksOne correct option
  1. A

    Local minimum

  2. B

    Local maximum

  3. C

    Saddle point

  4. D

    Neither minimum nor maximum

Show answer

Correct answer

  • C

    Saddle point

Question 11 on its own page

Question 12

+4 marksNumerical answer

A GAN Generator has the following fully connected architecture:
Input: latent noise vector of dimension 100 Hidden Layer 1: 50 units Hidden Layer 2: 50 units Output Layer: generates a 4 × 4 image (i.e., output dimension 16) Ignore all biases (count weights only)
How many total parameters does the Generator have?

Show answer

Correct answer: 8300

Question 12 on its own page

Question 13

+5 marksOne correct option
  1. A

    Perplexity: Low BLEU: High FID: Low IS: High CLIP: High

  2. B

    Perplexity: High BLEU: High FID: High IS: Low CLIP: High

  3. C

    Perplexity: Low BLEU: Low FID: High IS: Low CLIP: Low

  4. D

    Perplexity: High BLEU: Low FID: Low IS: High CLIP: Low

Show answer

Correct answer

  • A

    Perplexity: Low BLEU: High FID: Low IS: High CLIP: High

Question 13 on its own page

Question 14

+2 marksOne correct option
  1. A

    1.0

  2. B

    0.75

  3. C

    0.5

  4. D

    0.6

Show answer

Correct answer

  • B

    0.75

Question 14 on its own page

Question 15

+3 marksOne correct option
  1. A

    | Q | Decoder | | K | Encoder | | V | Encoder |

  2. B

    | Q | Encoder | | K | Decoder | | V | Decoder |

  3. C

    | Q | Decoder | | K | Decoder | | V | Encoder |

  4. D

    | Q | Encoder | | K | Encoder | | V | Decoder |

Show answer

Correct answer

  • A

    | Q | Decoder | | K | Encoder | | V | Encoder |

Question 15 on its own page

Question 16

+3 marksOne correct option

We train an encoder-decoder model.
The encoder has the following projection matrices:
Q: 5×4
K: 6×4
The decoder has the following projection matrices: Q: 6×4
K: 3×4
What will be the dimension of the attention matrix in the cross-attention layer?

  1. A

    5×3

  2. B

    6×6

  3. C

    5×6

  4. D

    6×3

Show answer

Correct answer

  • B

    6×6

Question 16 on its own page

Question 17

+3 marksOne correct option

You are building a machine translation system (English to Hindi) using an RNN-based Encoder-Decoder architecture. Your junior developer makes the following claims about the architecture:
Claim 1: "The encoder and decoder must have the same number of parameters (same hidden size, same number of layers)." Claim 2: "The encoder and decoder share the same weights during training to ensure consistency."
Which of the following statements is correct?

  1. A

    Both claims are TRUE - encoder and decoder must have identical architecture and share weights

  2. B

    Claim 1 is TRUE, Claim 2 is FALSE - they can have the same architecture but use separate, independent weights

  3. C

    Claim 1 is FALSE, Claim 2 is TRUE - they can have different architectures but must share weights

  4. D

    Both claims are FALSE - encoder and decoder can have different architectures (different hidden sizes, layers) and use separate, independent weights.

Show answer

Correct answer

  • D

    Both claims are FALSE - encoder and decoder can have different architectures (different hidden sizes, layers) and use separate, independent weights.

Question 17 on its own page

Question 18

+3 marksOne correct option
  1. A

    200

  2. B

    5

  3. C

    1000

  4. D

    205

Show answer

Correct answer

  • C

    1000

Question 18 on its own page

Question 19

+4 marksOne correct option
  1. A

    Because the formula only works for even values of i.

  2. B

    Because each i generates two dimensions (one sine and one cosine), so using (i = 0 to d−1) would produce 2d dimensions instead of d.

  3. C
  4. D

    Because cosine terms replace the sine terms for half the dimensions.

Show answer

Correct answer

  • B

    Because each i generates two dimensions (one sine and one cosine), so using (i = 0 to d−1) would produce 2d dimensions instead of d.

Question 19 on its own page

Question 20

+4 marksOne correct option
  1. A

    self.layer_norm(attention_output)

  2. B

    self.layer_norm(x + attention_output)

  3. C

    x + self.layer_norm(attention_output)

  4. D

    self.layer_norm(x) + attention_output

Show answer

Correct answer

  • B

    self.layer_norm(x + attention_output)

Question 20 on its own page

Question 21

+5 marksNumerical answer
Show answer

Correct answer: 0.5 (accepted within ±0.1)

Question 21 on its own page

Question 22

+4 marksOne correct option
  1. A

    81.2

  2. B

    75.3

  3. C

    68.1

  4. D

    40.9

Show answer

Correct answer

  • A

    81.2

Question 22 on its own page

Question 23

+4 marksOne correct option

Consider the following true and model distributions:
True distribution (P): Cat = 0.6 Dog = 0.3 Horse = 0.1
Model distribution (Q): Cat = 0.5 Dog = 0.4 Horse = 0.1
Which of the following statements about KL divergence is TRUE?

  1. A
  2. B

    KL divergence can be negative if Q assigns higher probability than P.

  3. C

    KL divergence is always non-negative, and equals 0 only if P and Q are identical.

  4. D

    KL divergence does not depend on the probability values of Q.

Show answer

Correct answer

  • C

    KL divergence is always non-negative, and equals 0 only if P and Q are identical.

Question 23 on its own page

Question 24

+2 marksOne correct option
  1. A

    Prevents repeated chunks

  2. B

    Ensures continuity of information between chunks

  3. C

    Reduces the total number of chunks

  4. D

    Removes unwanted tokens during splitting

Show answer

Correct answer

  • B

    Ensures continuity of information between chunks

Question 24 on its own page

Question 25

+2 marksOne correct option
  1. A

    A retriever that always returns 4 random chunks

  2. B

    A retriever that filters chunks with metadata length < 4

  3. C

    A retriever that searches only in the first 4 chunks

  4. D

    A retriever that returns the top-4 most similar chunks

Show answer

Correct answer

  • D

    A retriever that returns the top-4 most similar chunks

Question 25 on its own page

Question 26

+3 marksOne correct option
  1. A

    context_length,vocab_size,embedding_dimension

  2. B

    vocab_size,vocab_size,embedding_dimension

  3. C

    embedding_dimension,context_length,vocab_size

  4. D

    vocab_size,context_length,embedding_dimension

Show answer

Correct answer

  • D

    vocab_size,context_length,embedding_dimension

Question 26 on its own page

Question 27

+2 marksNumerical answer

Suppose there is RAG system containing a 100 page document with 10000 lines in total and it is segmented into 10 chunks, how many semantic search matches will be performed when a query enters the RAG system?

Show answer

Correct answer: 10

Question 27 on its own page

Question 28

+2 marksOne correct option
  1. A

    A list where each element is the text of one page

  2. B

    A single long string containing the text of all pages combined, with no separators added

  3. C

    Separate blocks of text with automatic page breaks inserted between pages

  4. D

    A dictionary mapping page numbers to extracted page text

Show answer

Correct answer

  • B

    A single long string containing the text of all pages combined, with no separators added

Question 28 on its own page

Question 29

+2 marksOne correct option
  1. A

    They must have the same weights (weight sharing)

  2. B

    They are independent learned transformations of the input

  3. C

    Key and Query must have the same weights, Value is independent

  4. D

    They implement dropout regularization

Show answer

Correct answer

  • B

    They are independent learned transformations of the input

Question 29 on its own page

Question 30

+4 marksOne correct option
  1. A

    Only retriever and format_docs

  2. B

    retriever | format_docs runs in parallel with RunnablePassthrough()

  3. C

    prompt_template runs in parallel with llm

  4. D

    Nothing runs in parallel; the entire chain is sequential

Show answer

Correct answer

  • B

    retriever | format_docs runs in parallel with RunnablePassthrough()

Question 30 on its own page