Responsible and Safe AI Systems Nptel Week 2 Answers
Are you looking for Responsible and Safe AI Systems Nptel Week 1 Answers ? All weeks solutions of this Swayam course are available here.
Table of Contents
Responsible and Safe AI Systems Nptel Week 2 Answers (July-Dec 2025)
Course link: Click here to visit course on Nptel Website
Question 1. Which of the following is true?
a) Extremistan has thin tails while Mediocristan has long tails
b) Mediocristan distributions are harder to predict than Extremistan
c) In Extremistan, the total is determined by a few large events with tyranny of the accidental
d) Extremistan has mild randomness while Mediocristan has wild randomness
Question 2. What is the key difference between covariate shift and concept shift in distribution shifts?
a) Covariate shift changes P(y|x) while concept shift changes P(x)
b) Covariate shift changes P(x) while P(y|x) remains constant, concept shift changes P(y|x) while P(x) remains constant
c) Both change P(x) and P(y|x) simultaneously
d) Covariate shift affects labels while concept shift affects features
Question 3. In the AugMix methodology, what is the primary advantage over uncontrolled random augmentations?
a) It uses skip connections to keep images recognizable while applying diverse augmentations
b) It requires less computational power
c) It only applies single augmentations instead of multiple
d) It focuses on geometric transformations only
Question 4. In the context of adversarial attacks, what does “transferability” specifically refer to?
a) The ability to transfer attacks from one domain to another
b) The ability to transfer defenses across different architectures
c) The ability to convert white-box attacks to black-box attacks
d) The ability of adversarial examples crafted for one model to work on other models
Question 5. Black Swan lies in which of the following categories?
a) Known Knowns
b) Known Unknowns
c) Unknown Knowns
d) Unknown Unknowns
Question 6. Which of the following are valid approaches for defending against adversarial attacks? (Select all that apply)
a) Data augmentation techniques
b) Adversarial training using adversarial examples during training
c) Using more data and larger models
d) Reducing model complexity to avoid overfitting
e) Adversarial pretraining on larger datasets like ImageNet
Question 7. In the RLHF optimization objective, why is a KL-divergence penalty term added to the reward maximization?
a) To prevent the model from generating repetitive outputs
b) To ensure the model stays close to the original pretrained model
c) To improve the computational efficiency of the training process
d) To increase the diversity of generated samples
Question 8. What does “reward hacking” specifically refer to in the context of RLHF?
a) Humans providing incorrect feedback to manipulate the system
b) External attackers compromising the reward model
c) The reward model overfitting to the training data
d) The model finding ways to maximize the reward function without achieving the intended behavior
Question 9. Identify the equations that can lead to a long-tailed distribution.
a) Idea × student × resources × time
b) Idea × student + resources × time
c) Idea + student + resource + time
d) Idea – student × resource – time
Question 10. What is the primary advantage of using pairwise comparisons over direct scalar ratings in human feedback collection?
a) Pairwise comparisons are faster to collect
b) They require fewer human annotators
c) Human judgments are noisy and miscalibrated, but pairwise comparisons are more reliable
d) They provide more granular feedback information
Question 11. What is a major challenge with using a single reward function in RLHF?
a) It is computationally expensive to optimize
b) It cannot represent a diverse society of humans
c) It requires too much training data
d) It is unstable during training
Question 12. What is Direct Preference Optimization (DPO) equivalent to in the context of RLHF component removal?
a) RLHF – Human Feedback
b) RLHF – Reward Model
c) RLHF – RL
d) RLHF – Policy Optimization
These are Responsible and Safe AI Systems Nptel Week 2 Answers