CS 7643 DEEP LEARNING QUIZ 1 |
QUESTIONS AND ANSWERS | 2026/27
UPDATE | 100% CORRECT - GT.
148 Questions with Answers and Detailed Rationales
100 PERCENT GUARANTEED PASS
INSTANT DOWNLOAD ANSWERS INCLUDED
IMPORTANCE OF THIS DOCUMENT
This comprehensive examination preparation guide has been meticulously developed to help you succeed in the
CS 7643 DEEP LEARNING QUIZ 1 | QUESTIONS AND ANSWERS | 2026/27 UPDATE | 100% CORRECT - GT..
It contains 148 carefully selected questions that reflect the most current exam content and testing strategies.
Each question is accompanied by a correct answer and a detailed rationale that explains the underlying
pathophysiology, pharmacology, or clinical reasoning.
Self-Assessment – Test your knowledge and Exam Preparation – Familiarize yourself with the
identify areas requiring further question format and content
study areas
Concept Reinforcement – Deepen your Confidence Building – Develop test-taking
understanding through strategies and reduce
evidence-based exam anxiety
rationales
Time Management – Practice answering
questions under simulated
exam conditions
Review Summary 148 Questions
Foundations - Application - CS 7643 DEEP Learning 1 AND 2026/27 Update 100 Correct - GT DEEP
Learning CS 7643 Graduate M S /ph D Level R1 University
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Linear Classifiers AND 1-25 Layer, Softmax, Stride, Padding, Input
Gradient Descent
Neural Networks AND 26-50 Training, Accuracy, Layer, Input, Output
Backpropagation
Regularization AND 51-75 Training, Network, Layer, Input, Output
Optimization
Convolutional Neural 76-100 Network, Accuracy, Output, Training, Gradients
Networks
Recurrent Neural Networks 101-125 Training, Accuracy, Layer, Model, Networks
Attention AND Transformers 126-148 Training, Layer, Output, Network, Through
TOTAL 148 All questions include answers and detailed rationales
,Section A - Linear Classifiers AND Gradient Descent
Q1.
A 3-layer MLP uses ReLU activations. During backpropagation, you observe that gradients
for early layers are exactly zero for most samples. Which initialization is most likely
responsible?
A. Xavier/Glorot initialization B. All weights initialized to zero
C. He initialization D. Small random normal with std 0.01
Correct: B - All weights initialized to zero
Rationale:Zero initialization makes all neurons in a layer compute identical outputs and
gradients, and with ReLU, if pre-activations are zero, gradients are zero, causing symmetry
and vanishing gradients. Xavier/He and small random init break symmetry and allow learning.
Thus zero init is the direct cause of exactly zero gradients.
Why the other answers are wrong:
A. Xavier initialization is designed to maintain variance and does not cause zero gradients.
C. He initialization is appropriate for ReLU and prevents vanishing gradients.
D. Small random initialization breaks symmetry and yields non-zero gradients, though it may
slow learning.
Reference: Goodfellow et al., Deep Learning, MIT Press, 2016, Ch. 6
Q2.
Which property distinguishes Adam from RMSProp?
A. Adam uses momentum on the first B. Adam uses a fixed learning rate.
moment and bias correction.
C. Adam normalizes gradients by their L2 D. Adam requires second-order derivatives.
norm.
Correct: A - Adam uses momentum on the first moment and bias correction.
Rationale:Adam incorporates both first-moment (momentum) and second-moment estimates
with bias correction, while RMSProp only uses the second moment. Adam does not use fixed
learning rate (it adapts per-parameter), does not normalize by L2 norm, and does not use
second-order derivatives.
Why the other answers are wrong:
B. Adam uses adaptive per-parameter learning rates, not a fixed one.
C. Adam scales by the square root of the second moment, not the L2 norm of the gradient.
D. Adam is a first-order method; it does not compute Hessians.
Reference: Kingma & Ba, Adam: A Method for Stochastic Optimization, ICLR 2015
Page 3
, Section A - Linear Classifiers AND Gradient Descent
Q3.
In a CNN, a 5x5 convolution with stride 1 and no padding is applied to a 32x32x3 input
with 64 filters. What is the output spatial dimension?
A. 28x28 B. 30x30
C. 32x32 D. 26x26
Correct: A - 28x28
Rationale:Output size = (W - F + 2P)/S + 1 = (32 - 5 + 0)/1 + 1 = 28. Thus 28x28. The other
options result from incorrect padding or stride assumptions.
Why the other answers are wrong:
B. 30x30 would result if padding were 1, but padding is 0.
C. 32x32 would require padding to preserve size, which is not used here.
D. 26x26 would result from a larger filter or different stride.
Reference: CS231n Convolutional Neural Networks for Visual Recognition, Stanford, 2025
Q4.
A recurrent neural network suffers from exploding gradients. Which technique most
directly addresses this during training?
A. Dropout on recurrent connections B. Gradient clipping
C. Batch normalization D. Weight decay
Correct: B - Gradient clipping
Rationale:Gradient clipping rescales gradients when their norm exceeds a threshold, directly
preventing exploding gradients. Dropout, batch norm, and weight decay help with
generalization or vanishing gradients but do not directly cap gradient magnitude.
Why the other answers are wrong:
A. Dropout on recurrent connections can harm learning and does not prevent gradient
explosion.
C. Batch normalization stabilizes activations but does not directly clip gradients.
D. Weight decay penalizes large weights but does not guarantee gradient clipping.
Reference: Pascanu et al., On the difficulty of training recurrent neural networks, ICML 2013
Q5.
Which statement about Batch Normalization (BN) during inference is correct?
A. BN uses the current mini-batch statistics B. BN uses running averages of mean and
at test time. variance estimated during training.
Page 4
QUESTIONS AND ANSWERS | 2026/27
UPDATE | 100% CORRECT - GT.
148 Questions with Answers and Detailed Rationales
100 PERCENT GUARANTEED PASS
INSTANT DOWNLOAD ANSWERS INCLUDED
IMPORTANCE OF THIS DOCUMENT
This comprehensive examination preparation guide has been meticulously developed to help you succeed in the
CS 7643 DEEP LEARNING QUIZ 1 | QUESTIONS AND ANSWERS | 2026/27 UPDATE | 100% CORRECT - GT..
It contains 148 carefully selected questions that reflect the most current exam content and testing strategies.
Each question is accompanied by a correct answer and a detailed rationale that explains the underlying
pathophysiology, pharmacology, or clinical reasoning.
Self-Assessment – Test your knowledge and Exam Preparation – Familiarize yourself with the
identify areas requiring further question format and content
study areas
Concept Reinforcement – Deepen your Confidence Building – Develop test-taking
understanding through strategies and reduce
evidence-based exam anxiety
rationales
Time Management – Practice answering
questions under simulated
exam conditions
Review Summary 148 Questions
Foundations - Application - CS 7643 DEEP Learning 1 AND 2026/27 Update 100 Correct - GT DEEP
Learning CS 7643 Graduate M S /ph D Level R1 University
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Linear Classifiers AND 1-25 Layer, Softmax, Stride, Padding, Input
Gradient Descent
Neural Networks AND 26-50 Training, Accuracy, Layer, Input, Output
Backpropagation
Regularization AND 51-75 Training, Network, Layer, Input, Output
Optimization
Convolutional Neural 76-100 Network, Accuracy, Output, Training, Gradients
Networks
Recurrent Neural Networks 101-125 Training, Accuracy, Layer, Model, Networks
Attention AND Transformers 126-148 Training, Layer, Output, Network, Through
TOTAL 148 All questions include answers and detailed rationales
,Section A - Linear Classifiers AND Gradient Descent
Q1.
A 3-layer MLP uses ReLU activations. During backpropagation, you observe that gradients
for early layers are exactly zero for most samples. Which initialization is most likely
responsible?
A. Xavier/Glorot initialization B. All weights initialized to zero
C. He initialization D. Small random normal with std 0.01
Correct: B - All weights initialized to zero
Rationale:Zero initialization makes all neurons in a layer compute identical outputs and
gradients, and with ReLU, if pre-activations are zero, gradients are zero, causing symmetry
and vanishing gradients. Xavier/He and small random init break symmetry and allow learning.
Thus zero init is the direct cause of exactly zero gradients.
Why the other answers are wrong:
A. Xavier initialization is designed to maintain variance and does not cause zero gradients.
C. He initialization is appropriate for ReLU and prevents vanishing gradients.
D. Small random initialization breaks symmetry and yields non-zero gradients, though it may
slow learning.
Reference: Goodfellow et al., Deep Learning, MIT Press, 2016, Ch. 6
Q2.
Which property distinguishes Adam from RMSProp?
A. Adam uses momentum on the first B. Adam uses a fixed learning rate.
moment and bias correction.
C. Adam normalizes gradients by their L2 D. Adam requires second-order derivatives.
norm.
Correct: A - Adam uses momentum on the first moment and bias correction.
Rationale:Adam incorporates both first-moment (momentum) and second-moment estimates
with bias correction, while RMSProp only uses the second moment. Adam does not use fixed
learning rate (it adapts per-parameter), does not normalize by L2 norm, and does not use
second-order derivatives.
Why the other answers are wrong:
B. Adam uses adaptive per-parameter learning rates, not a fixed one.
C. Adam scales by the square root of the second moment, not the L2 norm of the gradient.
D. Adam is a first-order method; it does not compute Hessians.
Reference: Kingma & Ba, Adam: A Method for Stochastic Optimization, ICLR 2015
Page 3
, Section A - Linear Classifiers AND Gradient Descent
Q3.
In a CNN, a 5x5 convolution with stride 1 and no padding is applied to a 32x32x3 input
with 64 filters. What is the output spatial dimension?
A. 28x28 B. 30x30
C. 32x32 D. 26x26
Correct: A - 28x28
Rationale:Output size = (W - F + 2P)/S + 1 = (32 - 5 + 0)/1 + 1 = 28. Thus 28x28. The other
options result from incorrect padding or stride assumptions.
Why the other answers are wrong:
B. 30x30 would result if padding were 1, but padding is 0.
C. 32x32 would require padding to preserve size, which is not used here.
D. 26x26 would result from a larger filter or different stride.
Reference: CS231n Convolutional Neural Networks for Visual Recognition, Stanford, 2025
Q4.
A recurrent neural network suffers from exploding gradients. Which technique most
directly addresses this during training?
A. Dropout on recurrent connections B. Gradient clipping
C. Batch normalization D. Weight decay
Correct: B - Gradient clipping
Rationale:Gradient clipping rescales gradients when their norm exceeds a threshold, directly
preventing exploding gradients. Dropout, batch norm, and weight decay help with
generalization or vanishing gradients but do not directly cap gradient magnitude.
Why the other answers are wrong:
A. Dropout on recurrent connections can harm learning and does not prevent gradient
explosion.
C. Batch normalization stabilizes activations but does not directly clip gradients.
D. Weight decay penalizes large weights but does not guarantee gradient clipping.
Reference: Pascanu et al., On the difficulty of training recurrent neural networks, ICML 2013
Q5.
Which statement about Batch Normalization (BN) during inference is correct?
A. BN uses the current mini-batch statistics B. BN uses running averages of mean and
at test time. variance estimated during training.
Page 4
