IT 621 Advanced AI & Machine Learning
Comprehensive Final Exam (Qns & Ans)
2025
Question 1 (Multiple Choice)
Question:
Which optimization algorithm improves generalization in deep
neural networks by decoupling weight decay from adaptive
learning rate updates?
A) Adam
B) RMSProp
C) AdamW
D) Adagrad
Correct ANS:
C) AdamW
©2025
, Rationale:
AdamW modifies the Adam optimizer by decoupling the weight
decay term from the gradient-based update. This leads to more
effective regularization and improved generalization in deep
learning cases, especially when training complex network
architectures.
---
Question 2 (Fill in the Blank)
Question:
The technique of incorporating specially crafted perturbations
during training to improve a model’s robustness against
adversarial inputs is known as ________ training .
Correct ANS:
adversarial
Rationale:
Adversarial training integrates adversarial examples—inputs
perturbed in a purposeful manner—into the training process. This
helps the model learn to resist small but malicious modifications,
thereby improving its overall robustness against adversarial
attacks.
©2025
,---
Question 3 (True/False)
Question:
True/False: In deep Q-learning, techniques such as experience
replay and separate target networks are critical stabilizers when
using function approximation with neural networks.
Correct ANS:
True
Rationale:
Deep Q-Learning can become unstable when using neural
networks to approximate Q-values. Techniques like experience
replay (which breaks correlation between sequential samples) and
separate target networks (which reduce bias and rapid policy
shifts) are essential to ensure stable and reliable learning.
---
Question 4 (Multiple Response)
©2025
, Question:
Select all techniques that help address gradient-related issues,
such as vanishing and exploding gradients, during the training of
deep neural networks:
A) Batch normalization
B) Skip connections
C) Gradient clipping
D) Data augmentation
E) Appropriate weight initialization
Correct ANS:
A, B, C, E
Rationale:
Batch normalization, skip (residual) connections, gradient
clipping, and careful weight initialization all contribute to
stabilizing gradients during training. Data augmentation enhances
generalization but does not directly mitigate gradient issues.
---
Question 5 (Multiple Choice)
©2025
Comprehensive Final Exam (Qns & Ans)
2025
Question 1 (Multiple Choice)
Question:
Which optimization algorithm improves generalization in deep
neural networks by decoupling weight decay from adaptive
learning rate updates?
A) Adam
B) RMSProp
C) AdamW
D) Adagrad
Correct ANS:
C) AdamW
©2025
, Rationale:
AdamW modifies the Adam optimizer by decoupling the weight
decay term from the gradient-based update. This leads to more
effective regularization and improved generalization in deep
learning cases, especially when training complex network
architectures.
---
Question 2 (Fill in the Blank)
Question:
The technique of incorporating specially crafted perturbations
during training to improve a model’s robustness against
adversarial inputs is known as ________ training .
Correct ANS:
adversarial
Rationale:
Adversarial training integrates adversarial examples—inputs
perturbed in a purposeful manner—into the training process. This
helps the model learn to resist small but malicious modifications,
thereby improving its overall robustness against adversarial
attacks.
©2025
,---
Question 3 (True/False)
Question:
True/False: In deep Q-learning, techniques such as experience
replay and separate target networks are critical stabilizers when
using function approximation with neural networks.
Correct ANS:
True
Rationale:
Deep Q-Learning can become unstable when using neural
networks to approximate Q-values. Techniques like experience
replay (which breaks correlation between sequential samples) and
separate target networks (which reduce bias and rapid policy
shifts) are essential to ensure stable and reliable learning.
---
Question 4 (Multiple Response)
©2025
, Question:
Select all techniques that help address gradient-related issues,
such as vanishing and exploding gradients, during the training of
deep neural networks:
A) Batch normalization
B) Skip connections
C) Gradient clipping
D) Data augmentation
E) Appropriate weight initialization
Correct ANS:
A, B, C, E
Rationale:
Batch normalization, skip (residual) connections, gradient
clipping, and careful weight initialization all contribute to
stabilizing gradients during training. Data augmentation enhances
generalization but does not directly mitigate gradient issues.
---
Question 5 (Multiple Choice)
©2025
