uiz Space

September 2025 term · Deep Learning for Computer Vision · BSDA5006

Deep Learning for Computer Vision End Term: 21 December 2025 (September 2025 term)

The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) End Term paper sat on 21 Dec 2025, in the September 2025 term: 19 questions for 19 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
19
Marks
19
Duration
180 min
Numerical
7
MCQ
8
MSQ
4
Take as mock testPractise with answers

Updated

Official paper: Deep Learning For Computer Vision 18 Dec 25 · No negative marking.

Question 1

+1 markNumerical answer
Show answer

Correct answer: 62

Question 1 on its own page

Question 2

+1 markNumerical answer
Show answer

Correct answer: 93 (accepted within ±1)

Question 2 on its own page

Question 3

+1 markOne correct option
  1. A

    Passing patch embeddings through the Transformer encoder

  2. B

    Linear projection of flattened image patches

  3. C

    Adding a class token

  4. D

    Layer normalization of patch embeddings

Show answer

Correct answer

  • B

    Linear projection of flattened image patches

Question 3 on its own page

Question 4

+1 markOne correct option
  1. A

    Flattens the softmax distribution, reducing gradient magnitude

  2. B

    sharpens the softmax distribution, increasing emphasis on hard negatives

  3. C

    has no effect on the loss landscape

  4. D

    makes image and text embeddings orthogonal

Show answer

Correct answer

  • B

    sharpens the softmax distribution, increasing emphasis on hard negatives

Question 4 on its own page

Question 5

+1 markOne correct option

In a Feature Pyramid Network (FPN), the primary purpose of the top-down pathway with lateral connections is to:

  1. A

    increase the receptive field at the lowest-resolution level

  2. B

    enrich high-resolution feature maps with strong semantics from deeper layers

  3. C

    perform non-maximum suppression at multiple scales

  4. D

    reduce the number of anchors required per level

Show answer

Correct answer

  • B

    enrich high-resolution feature maps with strong semantics from deeper layers

Question 5 on its own page

Question 6

+1 markOne correct option

Consider the following pyTorch code:

  1. A

    tensor([0.09, 0.24, 0.67])

  2. B

    tensor([0.09, 0.24, 0.66])

  3. C

    tensor([0.09, 0.23, 0.68])

  4. D

    tensor([0.10, 0.27, 0.63])

Show answer

Correct answer

  • A

    tensor([0.09, 0.24, 0.67])

Question 6 on its own page

Question 7

+1 markOne correct option
  1. A

    Use the test set to select the best hyperparameters

  2. B

    Use k-fold cross-validation on the training data and keep the test set untouched for final re-porting

  3. C

    Tune on the training loss only; validation is unnecessary

  4. D

    Pick hyperparameters from a prior paper without checking performance

Show answer

Correct answer

  • B

    Use k-fold cross-validation on the training data and keep the test set untouched for final re-porting

Question 7 on its own page

Question 8

+1 markOne correct option

How is disentangled representation evaluated?:

  1. A

    Using Mutual Information Gap

  2. B

    Using mean Average Precision

  3. C

    KL divergence

  4. D

    Sparsity in the latent one-hot codes

Show answer

Correct answer

  • A

    Using Mutual Information Gap

Question 8 on its own page

Question 9

+1 markOne or more correct options

Which statements about generative models are true? (Select all that apply.)

Select all that apply.

  1. A

    VAEs optimize a lower bound on the data log-likelihood.

  2. B

    GANs perform explicit maximum likelihood estimation.

  3. C

    Diffusion models can trade off speed and quality via the number of sampling steps.

  4. D

    Autoregressive models factorize the joint distribution into a product of conditionals.

Show answer

Correct answers

  • A

    VAEs optimize a lower bound on the data log-likelihood.

  • C

    Diffusion models can trade off speed and quality via the number of sampling steps.

  • D

    Autoregressive models factorize the joint distribution into a product of conditionals.

Question 9 on its own page

Question 10

+1 markOne or more correct options

In classifier-free guidance for diffusion models, increasing the guidance scale γ tends to: (Select all that apply.)

Select all that apply.

  1. A

    improve fidelity to the conditioning signal

  2. B

    always reduce mode collapse to zero

  3. C

    reduce sample diversity at very high values

  4. D

    be equivalent to classifier guidance

Show answer

Correct answers

  • A

    improve fidelity to the conditioning signal

  • C

    reduce sample diversity at very high values

Question 10 on its own page

Question 11

+1 markOne or more correct options

Which statements are correct? (Select all that apply.)

Select all that apply.

  1. A

    A linear autoencoder with MSE and no activation learns a subspace equivalent to PCA (up to rotation) when the hidden dimension < input dimension.

  2. B

    Global attention assigns alignment weights to all encoder time steps for each decoder step.

  3. C

    Positional encodings convey order information that self-attention lacks.

  4. D

    it is impossible to generate multiple, semantically similar captions with different styles for the same image.

Show answer

Correct answers

  • A

    A linear autoencoder with MSE and no activation learns a subspace equivalent to PCA (up to rotation) when the hidden dimension < input dimension.

  • B

    Global attention assigns alignment weights to all encoder time steps for each decoder step.

  • C

    Positional encodings convey order information that self-attention lacks.

Question 11 on its own page

Question 12

+1 markNumerical answer
Show answer

Correct answer: 27

Question 12 on its own page

Question 13

+1 markNumerical answer
Show answer

Correct answer: 50

Question 13 on its own page

Question 14

+1 markNumerical answer
Show answer

Correct answer: 156

Question 14 on its own page

Question 15

+1 markOne correct option

Mixup data augmentation primarily:

  1. A

    reduces label noise by hardening labels

  2. B

    creates convex combinations of inputs and labels to encourage linear behavior between classes

  3. C

    prunes redundant training samples

  4. D

    increases input resolution without changing model capacity

Show answer

Correct answer

  • B

    creates convex combinations of inputs and labels to encourage linear behavior between classes

Question 15 on its own page

Question 16

+1 markNumerical answer
Show answer

Correct answer: 0.645 (accepted within ±0.045)

Question 16 on its own page

Question 17

+1 markOne correct option

During training, Batch Normalization uses:

  1. A

    Running (population) statistics only

  2. B

    batch statistics for normalization and updates running estimates for inference

  3. C

    neither batch nor running statistics

  4. D

    group-wise normalization identical to GroupNorm

Show answer

Correct answer

  • B

    batch statistics for normalization and updates running estimates for inference

Question 17 on its own page

Question 18

+1 markOne or more correct options

Which techniques help mitigate exploding/vanishing gradients in RNNs? (Select all that apply.)

Select all that apply.

  1. A

    Gradient clipping

  2. B

    Orthogonal or identity initialization of recurrent weights

  3. C

    Using LSTM/GRU cells

  4. D

    Randomly reversing input sequences as augmentation

Show answer

Correct answers

  • A

    Gradient clipping

  • B

    Orthogonal or identity initialization of recurrent weights

  • C

    Using LSTM/GRU cells

Question 18 on its own pageAlso asked in Quiz 2 23 Nov 2025

Question 19

+1 markNumerical answer
Show answer

Correct answer: 6

Question 19 on its own page