The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) Quiz 2 paper sat on 24 Mar 2024, in the January 2024 term: 28 questions for 50 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.
Which of the following statements are true?
Tanh activation function can be represented as below in terms of sigmoid activation function:
tanh(x) = 2σ(2x) − 1, where σ = sigmoid activation function
The derivative of the Sigmoid activation function is symmetric around the origin.
Gradient of sigmoid neuron at saturation explodes
sigmoid activation is centered around 0.5 whereas tanh activation is centered around 0.
Correct answers
Tanh activation function can be represented as below in terms of sigmoid activation function:
tanh(x) = 2σ(2x) − 1, where σ = sigmoid activation function
sigmoid activation is centered around 0.5 whereas tanh activation is centered around 0.
Which of the below properties are not exhibited by the gradient-based explanation generation method, Grad-CAM? (Select all that apply)
Grad-CAM saliency map only computes gradients for important pixels while calculating weights, so pixels that contribute more to a specific class are considered
Grad-CAM can’t generate explanation directly from the given CNN and needs additional retraining
Grad-CAM is very reliable at identifying multiple instances of objects with occlusions
Grad-CAM can help localize objects without positional supervision
Correct answer
Grad-CAM saliency map only computes gradients for important pixels while calculating weights, so pixels that contribute more to a specific class are considered
Which of the following statements accurately describes the visualization techniques used in Convolutional Neural Networks (CNNs)? (Select all that apply)
Activation maximization methods identify input image region that maximally activates specific neurons in the network.
Filter visualization techniques display the learned weights of convolutional filters in the network.
Kernel visualization involves directly observing the values of convolutional filters during network training.
Principal Component Analysis (PCA) is commonly employed to reduce the dimensionality of feature maps for visualization purposes.
Correct answers
Activation maximization methods identify input image region that maximally activates specific neurons in the network.
Filter visualization techniques display the learned weights of convolutional filters in the network.
Which of the following is true regarding Hard Attention and Soft Attention?
Soft Attention is smooth and differentiable
Variance reduction techniques are used to train Hard Attention models
Soft Attention is computationally cheaper than Hard Attention when the source input is large
The inference (test time) overhead is low in Hard Attention when compared to Soft Attention models
Correct answers
Soft Attention is smooth and differentiable
Variance reduction techniques are used to train Hard Attention models
The inference (test time) overhead is low in Hard Attention when compared to Soft Attention models
Frechet Inception Distance (FID) is used for measuring the feature distance between real and generated images to evaluate the goodness of image generation in deep generative models such as GANs. Select the limitations of FID from the given below options:
(Select all possible correct options)
May not capture all features
Needs small sample size
Slow to run
None of these
Correct answers
May not capture all features
Slow to run
If x is an input feature vector and y is the corresponding output/label, which distributions are modeled in a vanilla Generative Adversarial Network (GAN)?
(Select all possible correct options)
p(x)
p(x|y)
p(y|x)
p(x, y)
Correct answers
p(x)
p(x, y)
What is the size of the feature map after applying two successive convolution operations with given parameters? Image size = 64 x 64, Kernel size = 6 x 6, Padding = 1 and Stride = 2. (In calculation, take floor(x) whenever x is non-integer. If the answer is FxF, write F in the answer).
Correct answer: 14
Correct answer: 1152
1→ vi, 2→ iii, 3→ v, 4 → i
1→ i, 2→ iv, 3 → iii, 4→ ii
1→ i, 2→ v, 3 → iii, 4→ iv
1→ iv, 2→ iii, 3→ ii, 4 → i
Correct answer
1→ iv, 2→ iii, 3→ ii, 4 → i
Which method is commonly used to explain Convolutional Neural Networks (CNNs) through visualization?
Feature Concatenation
Dimension Reduction
Gradient-weighted Class Activation Mapping (Grad-CAM)
Random Initialization
Correct answer
Gradient-weighted Class Activation Mapping (Grad-CAM)
1→ iv, 2→ iii, 3→ ii, 4 → i
1→ ii, 2→ iii, 3 → i, 4 →iv
1→ iv, 2→ iii, 3→ i, 4 → ii
1→ ii, 2→ iv, 3 → iii, 4→ i
Correct answer
1→ iv, 2→ iii, 3→ i, 4 → ii
Consider two 10×10 bounding boxes (one on the upper left and one of the lower right) in an image with an overlapping region of 7×7. The Intersection over Union (IoU) score between the two boxes is (choose the closest value):
21%
25%
33%
37%
Correct answer
33%
Which one of the following statements is true?
The goal of triplet loss is to ensure that the distance between the anchor and the negative sample is at least ’margin’ less than the distance between anchor and positive sample.
Face identification is a multi-class classification problem whereas face verification is a binary classification problem
Both the goal of triplet loss is to ensure that the distance between the anchor and the negative sample is at least ’margin’ less than the distance between anchor and positive sample & Face identification is a multi-class classification problem whereas face verification is a binary classification problem
None of these
Correct answer
Face identification is a multi-class classification problem whereas face verification is a binary classification problem
What are Haar-like features primarily used for in computer vision?
Object detection
Image segmentation
Texture analysis
Image denoising
Correct answer
Object detection
Correct answer
Sigmoid of alignment scores
Softmax of alignment scores
ReLU of alignment scores
None of these
Correct answer
Softmax of alignment scores
Which of the following is not a component of a Transformer network?
Encoder
Decoder
Positional Encoding
Recurrent Connections
Correct answer
Recurrent Connections
Consider the following assertion and the reason for the assertion below :
Now, pick one among the options below:
Both assertion and reason are true, and the reason is the correct explanation of assertion
Both assertion and reason are true, but the reason is not the correct explanation of assertion
Assertion is true, the reason is false
Both assertion and reason are false
Correct answer
Both assertion and reason are true, and the reason is the correct explanation of assertion
1→ iii, 2→ ii, 3 → vi, 4→ i
1→ v, 2→i, 3 → vi, 4→ iii
1→ ii, 2→ i, 3 → ii, 4 → iii
1→ i, 2 →ii, 3 → vi, 4→ iv
Correct answer
1→ v, 2→i, 3 → vi, 4→ iii
A four-dimensional input x = [3, 4, 1, 2] 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 [0.3, 0,−0.4, 0.5] and bias is 0.2.
Consider the corresponding activation functions and give the corresponding value of z. Based on the above data, answer the given subquestions.
Linear______________
Correct answer: 1.7
A four-dimensional input x = [3, 4, 1, 2] 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 [0.3, 0,−0.4, 0.5] and bias is 0.2.
Consider the corresponding activation functions and give the corresponding value of z. Based on the above data, answer the given subquestions.
Indicator function____________
Correct answer: 1
A four-dimensional input x = [3, 4, 1, 2] 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 [0.3, 0,−0.4, 0.5] and bias is 0.2.
Consider the corresponding activation functions and give the corresponding value of z. Based on the above data, answer the given subquestions.
Leaky-ReLU_____________
Correct answer: 1.7
Fill in the blanks for the given sub questions.
KL divergence is not a distance metric because KL(P||Q) is___________ to KL(Q||P).
NOTE: Enter the exact answer without any extra space in the beginning or at the end.
Correct answer: not equal
Fill in the blanks for the given sub questions.
Also KL divergence is always greater than equal to _____________
Correct answer: 0
Fill in the blanks for the given sub questions.
Suppose P := Bernoulli(0.9) and Q := Bernoulli(0.1), then KL(P||Q) = c × log(d).
The value for c is_________
Correct answer: 0.9
Fill in the blanks for the given sub questions.
Suppose P := Bernoulli(0.9) and Q := Bernoulli(0.1), then KL(P||Q) = c × log(d).
The value for d is_________
Correct answer: 8
While training a Generative Adversarial Network, a generator network G(.) and a discriminative network D(.) plays a zero-sum adversarial game.
Based on the above data, answer the given subquestions
The values of D(G(z)) during the initial phase of training_____________
Correct answer: 0
While training a Generative Adversarial Network, a generator network G(.) and a discriminative network D(.) plays a zero-sum adversarial game.
Based on the above data, answer the given subquestions
After successfully training the GAN__________
Correct answer: 0.5
