Description
1 Theory (15 marks)
1. (2 marks) Below is the modified objective function for a Generative Adversarial Network (GAN)
where a and b are the class labels for the generated images and real images. Given a fixed generator
G, derive the optimum discriminator D∗
(x):
minD[
1
2
Ex∼pdata(x)
[(D(x) − b)
2
] + 1
2
Ez∼pz(z)
[(D(G(z)) − a)
2
]]
2. (2 marks) Why does Inception Score (IS) fail in evaluating a GAN? What can be done to overcome
that failure?
3. (2 marks) Say your class has a facial recognition-based attendance system whose trained model is
inaccessible. What subcategory of attacks would you use to make everyone’s face to be classified as
yours? Substantiate your answer with valid reason(s).
4. (2 marks) In CIFAR10, you are given a sample from the automobile class and are asked to make
an existing trained model misclassify it as both bird and horse. Write down the appropriate PGD
attack equation for it.
1
5. (2 marks) Contrastive loss in Self-Supervised Learning helps in classifying similar samples together
and away from dissimilar samples. How should it be modified in order to give us the control of the
minimum distance between similar and dissimilar samples?
6. (2 marks) Consider the commonly used form of loss function for contrastive learning as below:
L = −E
log exp(score(f(x), f(x
+))/τ )
exp(score(f(x), f(x
+))/τ ) + PN−1
j=1 exp(score(f(x), f(x
−
j
))/τ )
where τ is the temperature hyperparameter. How does the temperature parameter in contrastive
learning affect the probability distribution? Substantiate your answer.
7. (1+1=2 marks) Consider the image colorization problem:
(a) Why is framing it as a classification problem is preferred over framing it as a regression problem?
(b) Why LAB color space is preferred over RGB?
8. (1 mark) If a newly proposed method claims to be 50-way-6-shot learning strategy, what is the class
count in novel classes and base classes? How many samples are there in each?
2 Programming (25 marks)
• The programming questions are shared in “Assignment 6.zip”. Please follow the instructions in the
notebook. Turn-in the notebook via Google Classroom once you finish your work.
• Marks breakdown is as follows:
– Question 1: 14 marks
– Question 2: 11 marks
2
