GEORGIA TECH CS 7643 DEEP LEARNING
QUIZ 1 PRACTICE | ANSWERS &
DETAILED RATIONALES | FALL 2026.
147 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
GEORGIA TECH CS 7643 DEEP LEARNING QUIZ 1 PRACTICE | ANSWERS & DETAILED RATIONALES |
FALL 2026.. It contains 147 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 147 Questions
Foundations - Application - Georgia TECH CS 7643 DEEP Learning 1 & Detailed Rationales FALL 2026
DEEP Learning CS 7643 Georgia TECH Omscs Graduate Ms/phd R1 University
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Linear Classifiers AND 1-25 Training, Padding, Layer, Input, Filters
Gradient Descent
Neural Networks AND 26-50 Training, Layer, Output, Cross-entropy, Input
Backpropagation
Regularization AND 51-75 Training, Network, Input, Output, Length
Optimization
Convolutional Neural 76-100 Class, Training, Accuracy, Output, Model
Networks
Recurrent Neural Networks 101-125 Training, Output, Layer, Validation, Activations
Attention AND Transformers 126-147 Network, Training, Layer, Model, Input
TOTAL 147 All questions include answers and detailed rationales
,Section A - Linear Classifiers AND Gradient Descent
Q1.
Let f(x) = ||Ax - b||_2^2 with A R^{m×n}. Which expression gives _x f in the
numerator-layout convention?
A. 2A^T(Ax - b) B. 2A(Ax - b)
C. 2(Ax - b) D. A^T A x - A^T b
Correct: A - 2A^T(Ax - b)
Rationale:Expanding gives x^T A^T A x - 2b^T A x + b^T b; differentiating yields 2A^T A x -
2A^T b = 2A^T(Ax - b). The transpose A^T is required so the result lives in R^n, matching x's
dimension. Option B has wrong dimensions, C drops A^T entirely, and D forgets the factor of
2 from the quadratic term.
Why the other answers are wrong:
B. A(Ax - b) has dimension m, not n, so it cannot be the gradient with respect to x R^n.
C. Omitting A^T ignores the chain rule through the linear map A.
D. Missing the factor of 2 that arises from differentiating the quadratic term x^T A^T A x.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, MIT Press, Ch. 2 (Linear Algebra) &
Ch. 4 (Numerical Computation).
Q2.
A network's loss is dominated by a single training example whose gradient norm is 10×
larger than all others. Which technique most directly addresses this without changing the
model architecture?
A. Gradient clipping by global norm B. Adding L2 weight decay
C. Switching from SGD to Adam D. Increasing the batch size to the full
dataset
Correct: A - Gradient clipping by global norm
Rationale:Global-norm gradient clipping rescales the entire gradient vector when its norm
exceeds a threshold, capping the influence of outlier examples and stabilizing updates. L2
decay penalizes weights, not gradient magnitude; Adam rescales per-parameter but still
propagates the outlier; full-batch training would still include the outlier's large gradient.
Why the other answers are wrong:
B. Weight decay shrinks parameters but does not cap the outlier's oversized gradient
contribution.
C. Adam normalizes per-parameter magnitudes but does not prevent a single example from
dominating the update direction.
Page 3
, Section A - Linear Classifiers AND Gradient Descent
D. Full-batch training still includes the outlier's gradient, so its relative influence is unchanged.
Reference: Pascanu, Mikolov & Bengio (2013). On the difficulty of training recurrent neural networks,
ICML.
Q3.
A 3-layer MLP uses ReLU activations initialized with all weights set to 0.5 and all biases to
0. What is the primary consequence during training?
A. Symmetry breaking fails; all hidden units B. Gradients vanish because ReLU
learn identical features saturates at 0
C. The network overfits because of small D. Batch normalization is required to make
initialization the network trainable
Correct: A - Symmetry breaking fails; all hidden units learn identical features
Rationale:With identical weights and biases, all neurons in a layer compute identical
pre-activations and receive identical gradients, so they remain symmetric and learn the same
features. ReLU does not saturate at 0 in the vanishing sense (that is sigmoid/tanh); small
constant init does not cause overfitting; and BN is not strictly required to break symmetry.
Why the other answers are wrong:
B. ReLU's zero region produces zero gradient only for negative inputs; the issue here is
symmetry, not saturation.
C. Small initialization is associated with slow learning, not overfitting.
D. Batch normalization helps optimization but is not the fundamental fix for identical-weight
symmetry.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, Ch. 6 & 8 (Backpropagation and
Optimization).
Q4.
For a 2D convolution with input 32×32×3, 64 filters of size 5×5, stride 1, and 'valid' padding
(no padding), what is the output spatial size and parameter count (excluding bias)?
A. 28×28×64, 4,800 parameters B. 28×28×64, 4,864 parameters
C. 32×32×64, 4,800 parameters D. 28×28×3, 4,800 parameters
Correct: A - 28×28×64, 4,800 parameters
Rationale:Output size is (32 - 5 + 1) = 28 per spatial dimension, with 64 channels. Each filter
has 5×5×3 = 75 weights, and 64 filters give 64 × 75 = 4,800 parameters (biases excluded).
Option B incorrectly includes 64 biases; C ignores the valid-padding reduction; D keeps the
input channel count.
Why the other answers are wrong:
B. 4,864 adds 64 bias terms, but the question excludes biases.
Page 4
QUIZ 1 PRACTICE | ANSWERS &
DETAILED RATIONALES | FALL 2026.
147 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
GEORGIA TECH CS 7643 DEEP LEARNING QUIZ 1 PRACTICE | ANSWERS & DETAILED RATIONALES |
FALL 2026.. It contains 147 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 147 Questions
Foundations - Application - Georgia TECH CS 7643 DEEP Learning 1 & Detailed Rationales FALL 2026
DEEP Learning CS 7643 Georgia TECH Omscs Graduate Ms/phd R1 University
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Linear Classifiers AND 1-25 Training, Padding, Layer, Input, Filters
Gradient Descent
Neural Networks AND 26-50 Training, Layer, Output, Cross-entropy, Input
Backpropagation
Regularization AND 51-75 Training, Network, Input, Output, Length
Optimization
Convolutional Neural 76-100 Class, Training, Accuracy, Output, Model
Networks
Recurrent Neural Networks 101-125 Training, Output, Layer, Validation, Activations
Attention AND Transformers 126-147 Network, Training, Layer, Model, Input
TOTAL 147 All questions include answers and detailed rationales
,Section A - Linear Classifiers AND Gradient Descent
Q1.
Let f(x) = ||Ax - b||_2^2 with A R^{m×n}. Which expression gives _x f in the
numerator-layout convention?
A. 2A^T(Ax - b) B. 2A(Ax - b)
C. 2(Ax - b) D. A^T A x - A^T b
Correct: A - 2A^T(Ax - b)
Rationale:Expanding gives x^T A^T A x - 2b^T A x + b^T b; differentiating yields 2A^T A x -
2A^T b = 2A^T(Ax - b). The transpose A^T is required so the result lives in R^n, matching x's
dimension. Option B has wrong dimensions, C drops A^T entirely, and D forgets the factor of
2 from the quadratic term.
Why the other answers are wrong:
B. A(Ax - b) has dimension m, not n, so it cannot be the gradient with respect to x R^n.
C. Omitting A^T ignores the chain rule through the linear map A.
D. Missing the factor of 2 that arises from differentiating the quadratic term x^T A^T A x.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, MIT Press, Ch. 2 (Linear Algebra) &
Ch. 4 (Numerical Computation).
Q2.
A network's loss is dominated by a single training example whose gradient norm is 10×
larger than all others. Which technique most directly addresses this without changing the
model architecture?
A. Gradient clipping by global norm B. Adding L2 weight decay
C. Switching from SGD to Adam D. Increasing the batch size to the full
dataset
Correct: A - Gradient clipping by global norm
Rationale:Global-norm gradient clipping rescales the entire gradient vector when its norm
exceeds a threshold, capping the influence of outlier examples and stabilizing updates. L2
decay penalizes weights, not gradient magnitude; Adam rescales per-parameter but still
propagates the outlier; full-batch training would still include the outlier's large gradient.
Why the other answers are wrong:
B. Weight decay shrinks parameters but does not cap the outlier's oversized gradient
contribution.
C. Adam normalizes per-parameter magnitudes but does not prevent a single example from
dominating the update direction.
Page 3
, Section A - Linear Classifiers AND Gradient Descent
D. Full-batch training still includes the outlier's gradient, so its relative influence is unchanged.
Reference: Pascanu, Mikolov & Bengio (2013). On the difficulty of training recurrent neural networks,
ICML.
Q3.
A 3-layer MLP uses ReLU activations initialized with all weights set to 0.5 and all biases to
0. What is the primary consequence during training?
A. Symmetry breaking fails; all hidden units B. Gradients vanish because ReLU
learn identical features saturates at 0
C. The network overfits because of small D. Batch normalization is required to make
initialization the network trainable
Correct: A - Symmetry breaking fails; all hidden units learn identical features
Rationale:With identical weights and biases, all neurons in a layer compute identical
pre-activations and receive identical gradients, so they remain symmetric and learn the same
features. ReLU does not saturate at 0 in the vanishing sense (that is sigmoid/tanh); small
constant init does not cause overfitting; and BN is not strictly required to break symmetry.
Why the other answers are wrong:
B. ReLU's zero region produces zero gradient only for negative inputs; the issue here is
symmetry, not saturation.
C. Small initialization is associated with slow learning, not overfitting.
D. Batch normalization helps optimization but is not the fundamental fix for identical-weight
symmetry.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, Ch. 6 & 8 (Backpropagation and
Optimization).
Q4.
For a 2D convolution with input 32×32×3, 64 filters of size 5×5, stride 1, and 'valid' padding
(no padding), what is the output spatial size and parameter count (excluding bias)?
A. 28×28×64, 4,800 parameters B. 28×28×64, 4,864 parameters
C. 32×32×64, 4,800 parameters D. 28×28×3, 4,800 parameters
Correct: A - 28×28×64, 4,800 parameters
Rationale:Output size is (32 - 5 + 1) = 28 per spatial dimension, with 64 channels. Each filter
has 5×5×3 = 75 weights, and 64 filters give 64 × 75 = 4,800 parameters (biases excluded).
Option B incorrectly includes 64 biases; C ignores the valid-padding reduction; D keeps the
input channel count.
Why the other answers are wrong:
B. 4,864 adds 64 bias terms, but the question excludes biases.
Page 4
