Quiz Space

January 2025 term · Deep Learning for Computer Vision · BSDA5006

Deep Learning for Computer Vision End Term: 13 April 2025, Set 1-4 (January 2025 term)

The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) End Term paper sat on 13 Apr 2025, in the January 2025 term, set 1-4: 47 questions for 66 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
47
Marks
66
Duration
180 min
MCQ
12
MSQ
5
Numerical
30
Take as mock testPractise with answers

Updated

Official paper: IIT M IMPROVEMENT FN EXAM QIM2 13 Apr · No negative marking.

Question 1

+2 marksOne correct option

Why does DETR typically exhibit poor performance in detecting small objects compared to larger ones? Choose the best answer in your opinion.

  1. A

    The CNN backbone used in DETR has a fixed receptive field that’s too large forsmall objects.

  2. B

    The global self-attention mechanism in DETR tends to dilute the signal fromsmall objects across all image locations.

  3. C

    DETR’s loss function actively discards all detections smaller than 50 × 50 pixels.

  4. D

    The object queries in DETR are programmed to ignore any object smaller than10% of the image size.

  5. E

    DETR’s encoder-decoder architecture was specifically designed to only detectobjects larger than a cat.

Show answer

Correct answer

  • B

    The global self-attention mechanism in DETR tends to dilute the signal fromsmall objects across all image locations.

Question 1 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 2

+2 marksOne correct option

Which one of the following statements is false?

  1. A

    Linear contrast stretching is a point operation.

  2. B

    Moving average is an example of a local operation.

  3. C

    Convolution in the spatial domain can be obtained through addition in thefrequency domain.

  4. D

    All of these.

Show answer

Correct answer

  • C

    Convolution in the spatial domain can be obtained through addition in thefrequency domain.

Question 2 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 3

+2 marksOne correct option

What is the correct order of operations for processing an image through a Vision Transformer (ViT)?

  1. A

    Image patching → Positional embedding → Linear projection of flattenedpatches → Transformer encoder→ Classification head

  2. B

    Image patching → Linear projection of flattened patches → Positionalembedding → Transformer encoder→ Classification head

  3. C

    Positional embedding → Image patching → Linear projection of flattenedpatches → Transformer encoder→ Classification head

  4. D

    Linear projection of Images → Image patching → Positional embedding →Transformer encoder → Classification head

  5. E

    Linear projection of Images → Image patching → Positional embedding →Transformer encoder → Transformer decoder→ Classification head

Show answer

Correct answer

  • B

    Image patching → Linear projection of flattened patches → Positionalembedding → Transformer encoder→ Classification head

Question 3 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 4

+2 marksOne correct option

Vector Quantized Variational Autoencoder (VQ-VAE) utilizes a discrete latent representation as opposed to continuous latent spaces used in traditional VAEs. One of the key components of a VQ-VAE is the codebook, which consists of a set of learnable vectors. What is the primary role of the codebook in VQ-VAE?

  1. A

    It performs the non-linear transformation of the input data to a higherdimensional space.

  2. B

    It regularizes the encoder by penalizing complex encodings.

  3. C

    It provides a finite set of vectors which the encoder’s outputs are mapped to,effectively quantizing the latent space.

  4. D

    It decodes the quantized vectors back into the reconstructed input space.

Show answer

Correct answer

  • C

    It provides a finite set of vectors which the encoder’s outputs are mapped to,effectively quantizing the latent space.

Question 4 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 5

+2 marksOne correct option

Given two normalized embeddings from CLIP, I = [0.2,−0.5, 0.3, 0.4] for an image and T = [−0.1, 0.6, −0.3,−0.7] for a text description, compute the CLIPScore(I, T).

  1. A

    -69%

  2. B

    0%

  3. C

    -30%

  4. D

    50%

Show answer

Correct answer

  • A

    -69%

Question 5 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 6

+2 marksOne correct option

Which one of the following statements regarding hyperparameter tuning is false?

  1. A

    Validation set is the best choice for tuning hyperparameters

  2. B

    Picking hyperparameters that perform the best on the training set leads tooverfitting

  3. C

    Tuning hyperparameters on the test set is an acceptable approach to find themost powerful model for generalization performance

  4. D

    Test set performance, with hyperparameters tuned on the validation set,would be a realistic measure of how the model will generalize to new data.

Show answer

Correct answer

  • C

    Tuning hyperparameters on the test set is an acceptable approach to find themost powerful model for generalization performance

Question 6 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 7

+2 marksOne correct option

Which one of the following statements is false? (Pick the most appropriate one.)

  1. A

    Attention mechanisms cannot be applied to the bidirectional RNN model

  2. B

    An image captioning network cannot be trained end-to-end even though weare using 2 different modalities to train the network

  3. C

    One of the key components in the vanilla transformer is the recurrentconnection that help them to deal with variable input length.

  4. D

    All of these.

Show answer

Correct answer

  • D

    All of these.

Question 7 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 8

+2 marksOne correct option
  1. A

    1→ iii, 2→ iv, 3→ i, 4 → ii

  2. B

    1→ iii, 2→ i, 3 → ii, 4 → v

  3. C

    1→ iii, 2→ iv, 3→ v, 4 → ii

  4. D

    1→ iv, 2→ iii, 3→ i, 4 → ii

Show answer

Correct answer

  • C

    1→ iii, 2→ iv, 3→ v, 4 → ii

Question 8 on its own page

Question 9

+2 marksOne correct option
  1. A

    1→ iii, 2→ iv, 3→ ii, 4 → i

  2. B

    1→ i, 2→ iv, 3 → iii, 4→ ii

  3. C

    1→ iv, 2→ iii, 3→ ii, 4 → i

  4. D

    1→ iii, 2→ i, 3 → ii, 4 →v

Show answer

Correct answer

  • D

    1→ iii, 2→ i, 3 → ii, 4 →v

Question 9 on its own page

Question 10

+2 marksOne correct option
  1. A

    1→ iv, 2→ v, 3→ i, 4 → ii

  2. B

    1→ iii, 2→ i, 3 → v, 4→ ii

  3. C

    1→ iv, 2→ iii, 3→ ii, 4 → i

  4. D

    1→ iii, 2→ i, 3 → ii, 4→ v

Show answer

Correct answer

  • A

    1→ iv, 2→ v, 3→ i, 4 → ii

Question 10 on its own page

Question 11

+2 marksOne correct option

In which one of the following applications would you use a one- to-many RNN architecture? Choose the most appropriate answer.

  1. A

    Input is a review of a movie and output is the sentiment of the given review

  2. B

    Input is an image and output is the textual description of the given image

  3. C

    Input is an image and output is the class to which the given image belongs

  4. D

    Input is a sentence in English language and output is the same sentencetranslated to Hindi language

Show answer

Correct answer

  • B

    Input is an image and output is the textual description of the given image

Question 11 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 12

+2 marksOne correct option

Why might Segment Anything (SAM) be particularly useful in data annotation tasks compared to traditional segmentation models?

  1. A

    It produces perfect segmentation masks.

  2. B

    It can adapt to segment any object, even those not seen during training

  3. C

    It requires less computational resources

  4. D

    It automatically labels all objects in an image without user input

Show answer

Correct answer

  • B

    It can adapt to segment any object, even those not seen during training

Question 12 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 31 Aug 2025

Question 13

+2 marksOne or more correct options

Which of the following techniques help control the exploding or vanishing gradient problem in recurrent neural networks?

Select all that apply.

  1. A

    Gradient clipping

  2. B

    Use ReLU activation funtion to control exploding gradient and sigmoidactivation to control vanishing gradients

  3. C

    Use Long Short-Term Memory (LSTM)

  4. D

    Use data augmentation methods (for example, reversing the input sequence)

Show answer

Correct answers

  • A

    Gradient clipping

  • C

    Use Long Short-Term Memory (LSTM)

  • D

    Use data augmentation methods (for example, reversing the input sequence)

Question 13 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 14

+2 marksOne or more correct options

Which of the following statements are true? (Select all possible correct options)

Select all that apply.

  1. A

    Generative adversarial networks (GANs) generate sharper images comparedto Variational AutoEncoders (VAE)

  2. B

    GAN is an example of an implicit density estimation model

  3. C

    Fully connected layers in mapping network of Style-GAN change thedimension of its input

  4. D

    The generator and discriminator are always trained together in a GAN

Show answer

Correct answers

  • A

    Generative adversarial networks (GANs) generate sharper images comparedto Variational AutoEncoders (VAE)

  • B

    GAN is an example of an implicit density estimation model

Question 14 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 15

+2 marksOne or more correct options

Which of the following statements are true ? (Select all possible correct options)

Select all that apply.

  1. A

    Autoencoder are equivalent to Principal Component Analysis (PCA) providedwe don’t use of non-linear activation functions

  2. B

    When using global attention on temporal data, alignment weights are learntfor encoder hidden representations for all time steps

  3. C

    Positional encoding is an important component of the transformerarchitecture as it conveys information about order in a given sequence

  4. D

    It is not possible to generate different captions for the same image that havesimilar meaning but different tone / style

  5. E

    Autoencoders can not be used for data compression as its input and outputdimensions are different

Show answer

Correct answers

  • A

    Autoencoder are equivalent to Principal Component Analysis (PCA) providedwe don’t use of non-linear activation functions

  • B

    When using global attention on temporal data, alignment weights are learntfor encoder hidden representations for all time steps

  • C

    Positional encoding is an important component of the transformerarchitecture as it conveys information about order in a given sequence

Question 15 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 16

+2 marksOne or more correct options

Which of the following statements about CLIP are TRUE? (Select ALL that apply)

Select all that apply.

  1. A

    CLIP uses a contrastive loss function to align text and image embeddings.

  2. B

    CLIP requires image labels for training.

  3. C

    CLIP is trained to maximize cosine similarity between embeddings of matchingtext-image pairs.

  4. D

    CLIP model is trained with pretext tasks such as rotation prediction andinpainting to learn effective feature representations.

  5. E

    To use the CLIP model for image classification, it is imperative to finetune it ondownstream datasets.

Show answer

Correct answers

  • A

    CLIP uses a contrastive loss function to align text and image embeddings.

  • C

    CLIP is trained to maximize cosine similarity between embeddings of matchingtext-image pairs.

Question 16 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 17

+2 marksOne or more correct options

Which of the following statements are true?(Select ALL that apply)

Select all that apply.

  1. A

    Weight change criterion is a method of ‘early stopping’ that checks whether ornot the error is dropping over epochs to decide whether to continue training or stop.

  2. B

    L1 norm tends to create more sparse weights than L2 norm.

  3. C

    During the training phase, for each iteration, Dropout ignores a randomfraction, p, of nodes, and accounts for it in the test phase by scaling down the activations by a factor of p.

  4. D

    A single McCulloch-Pitts neuron is capable of modeling AND, OR, XOR, NOR,and NAND functions.

Show answer

Correct answers

  • B

    L1 norm tends to create more sparse weights than L2 norm.

  • C

    During the training phase, for each iteration, Dropout ignores a randomfraction, p, of nodes, and accounts for it in the test phase by scaling down the activations by a factor of p.

Question 17 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 18

+2 marksNumerical answer
Show answer

Correct answer: 0

Question 18 on its own page

Question 19

+2 marksNumerical answer
Show answer

Correct answer: 30

Question 19 on its own page

Question 20

+2 marksNumerical answer
Show answer

Correct answer: -0.21 (accepted within ±0.005)

Question 20 on its own page

Question 21

+2 marksNumerical answer

Consider a reverse process in a diffusion model where the goal is to reconstruct the original data from the noise. If the model correctly reduces the variance of the noise by 0.02 in each reverse step, and starts with a noise variance of 1.0 at timestep T = 50, how many steps are required to reduce the noise variance to 0.1 _______________ ?

Show answer

Correct answer: 45

Question 21 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 22

+2 marksNumerical answer
Show answer

Correct answer: 3.75

Question 22 on its own page

Question 23

+2 marksNumerical answer
Show answer

Correct answer: 2

Question 23 on its own page

Question 24

+2 marksNumerical answer
Show answer

Correct answer: 100

Question 24 on its own page

Question 25

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: -2

Question 25 on its own page

Question 26

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Show answer

Correct answer: 5

Question 26 on its own page

Question 27

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Element 1: __________________

Show answer

Correct answer: 0.54

Question 27 on its own page

Question 28

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Element 2: __________________

Show answer

Correct answer: 0.25

Question 28 on its own page

Question 29

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Element 3: __________________

Show answer

Correct answer: 0.19

Question 29 on its own page

Question 30

+1 markNumerical answer

Based on the above data, answer the given subquestions.

Element 4: __________________

Show answer

Correct answer: 0.58

Question 30 on its own page

Question 31

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Self-Attention _____________

Show answer

Correct answer: 1

Question 31 on its own page

Question 32

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

BERT Text Encoder Bi Self Attention: ________________

Show answer

Correct answer: 5

Question 32 on its own page

Question 33

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Image-Text Matching Loss (LITM ): _______________

Show answer

Correct answer: 11

Question 33 on its own page

Question 34

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Image Grounded Text Encoder Bi Self Attention: _______________

Show answer

Correct answer: 7

Question 34 on its own page

Question 35

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Show answer

Correct answer: 8

Question 35 on its own page

Question 36

+0.55 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Image Grounded Text Encoder Cross Attention: ______________

Show answer

Correct answer: 4

Question 36 on its own page

Question 37

+0.54 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Image-Text Contrastive Loss (LIT C): ______________

Show answer

Correct answer: 10

Question 37 on its own page

Question 38

+0.54 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Show answer

Correct answer: 9

Question 38 on its own page

Question 39

+0.54 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Causal Self Attention: ________________

Show answer

Correct answer: 3

Question 39 on its own page

Question 40

+0.54 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Show answer

Correct answer: 2

Question 40 on its own page

Question 41

+0.54 marksNumerical answer

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:

Consider the BLIP model architecture with an input (image-text pair) as shown below:
For each of the given entities, enter the appropriate number from the figure that corresponds to it:
Based on the above data, answer the given subquestions.

Image Grounded Text Deocder Cross Attention: ________________

Show answer

Correct answer: 6

Question 41 on its own page

Question 42

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Sigmoid ____________

Show answer

Correct answer: 0.18

Question 42 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 43

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Linear ______________

Show answer

Correct answer: -1.5

Question 43 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 44

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Indicator Function _______________

Show answer

Correct answer: 0

Question 44 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 45

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Softplus _____________

Show answer

Correct answer: 0.20

Question 45 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 46

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Relu _____________

Show answer

Correct answer: 0

Question 46 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025

Question 47

+1 markNumerical answer

A four-dimensional input vector x = [1,−2, 3,−4] is passed to a hidden layer with a single neuron and an activation function a(.) to obtain z. Assume that the corresponding weights (from input to hidden layer) are [0.5,−0.3, 0.8, 1.2] and bias is −0.2. If the activation function a(.) is Sigmoid, Linear, Indicator function, Softplus, ReLU and Leaky- ReLU, what are the corresponding values of z? (Note that Indicator function returns 1 for a positive input, and 0 otherwise. Softplus is defined as: ln(1 + exp(z))).
Based on the above data, answer the given subquestions.

Leaky-Relu ______________

Show answer

Correct answer: -0.015

Question 47 on its own pageAlso asked in End Term 13 Apr 2025, End Term 13 Apr 2025, End Term 13 Apr 2025