GEORGIA TECH CS 7643 DEEP LEARNING QUIZ 2
COMPLETE PRACTICE & REVIEW BUNDLE |
ANSWERS & DETAILED RATIONALES | FALL 2026
150 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 2 COMPLETE PRACTICE & REVIEW BUNDLE | ANSWERS
& DETAILED RATIONALES | FALL 2026. It contains 150 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 150 Questions
Foundations - Application - Georgia TECH CS 7643 DEEP Learning 2 Complete & Review Bundle &
Detailed Rationales FALL 2026 DEEP Learning Optimization Regularization CNNS RNNS Attention AND
Generative Models Graduate Ms/phd-level Georgia TECH CS 7643
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Neural Networks AND 1-25 Attention, Training, Model, Single, Stride
Backpropagation
Convolutional Neural 26-50 Training, Model, Accuracy, Layer, DEEP Networks
Networks
Recurrent Neural Networks 51-75 Training, Layer, Directly, Gradients, Input
AND Sequence Modeling
Attention Mechanisms AND 76-100 Training, Accuracy, Model, Validation, Directly
Transformers
Regularization Optimization 101-125 Training, Accuracy, Model, Single, Stride
AND Training Techniques
Generative Models VAES 126-150 Training, Model, Layer, Input, Accuracy
GANS Diffusion Models
TOTAL 150 All questions include answers and detailed rationales
,Section A - Neural Networks AND Backpropagation
Q1.
During training of a deep MLP, you observe that the training loss decreases smoothly
while the validation loss begins to rise steadily after epoch 15. Which single intervention
most directly targets the underlying cause?
A. Increase the learning rate to escape a B. Add L2 weight decay and/or dropout to
sharp minimum. constrain model capacity.
C. Switch from Adam to SGD with D. Add batch normalization layers to
momentum to improve generalization. stabilize hidden activations.
Correct: B - Add L2 weight decay and/or dropout to constrain model capacity.
Rationale:A widening gap where validation loss rises while training loss falls is the classic
signature of overfitting, so capacity-controlling regularization (weight decay, dropout, early
stopping) directly addresses it. Learning-rate or optimizer changes do not reduce the model's
effective capacity, and batch norm primarily stabilizes optimization rather than limiting
overfitting.
Why the other answers are wrong:
A. Increasing the learning rate can worsen overfitting and destabilize training rather than
reduce the generalization gap.
C. Changing optimizer may alter convergence but does not directly constrain model capacity to
close the generalization gap.
D. Batch normalization aids optimization stability but is not a primary remedy for overfitting.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, MIT Press, Ch. 7 (Regularization).
Q2.
A 3x3 convolution with stride 1 and no padding is applied to a 32x32xC input. What is the
spatial output size, and how many parameters does the kernel have (excluding bias)?
A. 30x30, with 9*C parameters per output B. 32x32, with 9*C parameters per output
channel. channel.
C. 30x30, with 9*C*C parameters per output D. 16x16, with 9*C parameters per output
channel. channel.
Correct: A - 30x30, with 9*C parameters per output channel.
Rationale:Valid 3x3 convolution reduces each spatial dimension by 2 (32-3+1=30). Each
output channel uses a 3x3xC kernel, i.e., 9*C weights, independent of the number of output
channels. Options B and D miscompute spatial size, and C overcounts parameters by
multiplying by C again.
Page 3
, Section A - Neural Networks AND Backpropagation
Why the other answers are wrong:
B. No padding means the spatial size shrinks; 32x32 is only preserved with 'same' padding.
C. Parameters per output channel are 3x3xC, not 3x3xCxC; the extra C factor is incorrect.
D. 16x16 would require stride 2 or pooling, neither of which is used here.
Reference: Stanford CS231n Notes, Convolutional Neural Networks (spatial dimensions and parameter
counting).
Q3.
In a standard LSTM cell, which gate directly controls how much of the previous cell state
is retained versus overwritten at the current time step?
A. Input gate B. Output gate
C. Forget gate D. Reset gate
Correct: C - Forget gate
Rationale:The forget gate f_t multiplies the previous cell state C_{t-1}, determining how much
old memory persists. The input gate controls new candidate information, the output gate
modulates the exposed hidden state, and 'reset gate' belongs to GRUs, not LSTMs.
Why the other answers are wrong:
A. The input gate scales the new candidate memory, not the retention of the old cell state.
B. The output gate controls how much of the cell state is exposed as the hidden state h_t.
D. Reset gates are components of GRU cells, not standard LSTM cells.
Reference: Hochreiter & Schmidhuber (1997). Long Short-Term Memory, Neural Computation; Colah's
LSTM blog.
Q4.
Which statement best characterizes the inductive bias of the self-attention mechanism
compared to a convolutional layer?
A. Self-attention is translation-equivariant by B. Self-attention has no built-in locality or
construction, while convolution is translation bias and must learn positional
permutation-invariant. information, whereas convolution
hard-codes local weight sharing.
C. Self-attention enforces a fixed receptive D. Self-attention requires fewer parameters
field per layer equal to its window size, like than convolution for the same sequence
convolution. length.
Correct: B - Self-attention has no built-in locality or translation bias and must learn
positional information, whereas convolution hard-codes local weight sharing.
Page 4
COMPLETE PRACTICE & REVIEW BUNDLE |
ANSWERS & DETAILED RATIONALES | FALL 2026
150 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 2 COMPLETE PRACTICE & REVIEW BUNDLE | ANSWERS
& DETAILED RATIONALES | FALL 2026. It contains 150 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 150 Questions
Foundations - Application - Georgia TECH CS 7643 DEEP Learning 2 Complete & Review Bundle &
Detailed Rationales FALL 2026 DEEP Learning Optimization Regularization CNNS RNNS Attention AND
Generative Models Graduate Ms/phd-level Georgia TECH CS 7643
All answers with rationales
,Table of Contents
Content Area Questions Key Topics
Neural Networks AND 1-25 Attention, Training, Model, Single, Stride
Backpropagation
Convolutional Neural 26-50 Training, Model, Accuracy, Layer, DEEP Networks
Networks
Recurrent Neural Networks 51-75 Training, Layer, Directly, Gradients, Input
AND Sequence Modeling
Attention Mechanisms AND 76-100 Training, Accuracy, Model, Validation, Directly
Transformers
Regularization Optimization 101-125 Training, Accuracy, Model, Single, Stride
AND Training Techniques
Generative Models VAES 126-150 Training, Model, Layer, Input, Accuracy
GANS Diffusion Models
TOTAL 150 All questions include answers and detailed rationales
,Section A - Neural Networks AND Backpropagation
Q1.
During training of a deep MLP, you observe that the training loss decreases smoothly
while the validation loss begins to rise steadily after epoch 15. Which single intervention
most directly targets the underlying cause?
A. Increase the learning rate to escape a B. Add L2 weight decay and/or dropout to
sharp minimum. constrain model capacity.
C. Switch from Adam to SGD with D. Add batch normalization layers to
momentum to improve generalization. stabilize hidden activations.
Correct: B - Add L2 weight decay and/or dropout to constrain model capacity.
Rationale:A widening gap where validation loss rises while training loss falls is the classic
signature of overfitting, so capacity-controlling regularization (weight decay, dropout, early
stopping) directly addresses it. Learning-rate or optimizer changes do not reduce the model's
effective capacity, and batch norm primarily stabilizes optimization rather than limiting
overfitting.
Why the other answers are wrong:
A. Increasing the learning rate can worsen overfitting and destabilize training rather than
reduce the generalization gap.
C. Changing optimizer may alter convergence but does not directly constrain model capacity to
close the generalization gap.
D. Batch normalization aids optimization stability but is not a primary remedy for overfitting.
Reference: Goodfellow, Bengio & Courville (2016). Deep Learning, MIT Press, Ch. 7 (Regularization).
Q2.
A 3x3 convolution with stride 1 and no padding is applied to a 32x32xC input. What is the
spatial output size, and how many parameters does the kernel have (excluding bias)?
A. 30x30, with 9*C parameters per output B. 32x32, with 9*C parameters per output
channel. channel.
C. 30x30, with 9*C*C parameters per output D. 16x16, with 9*C parameters per output
channel. channel.
Correct: A - 30x30, with 9*C parameters per output channel.
Rationale:Valid 3x3 convolution reduces each spatial dimension by 2 (32-3+1=30). Each
output channel uses a 3x3xC kernel, i.e., 9*C weights, independent of the number of output
channels. Options B and D miscompute spatial size, and C overcounts parameters by
multiplying by C again.
Page 3
, Section A - Neural Networks AND Backpropagation
Why the other answers are wrong:
B. No padding means the spatial size shrinks; 32x32 is only preserved with 'same' padding.
C. Parameters per output channel are 3x3xC, not 3x3xCxC; the extra C factor is incorrect.
D. 16x16 would require stride 2 or pooling, neither of which is used here.
Reference: Stanford CS231n Notes, Convolutional Neural Networks (spatial dimensions and parameter
counting).
Q3.
In a standard LSTM cell, which gate directly controls how much of the previous cell state
is retained versus overwritten at the current time step?
A. Input gate B. Output gate
C. Forget gate D. Reset gate
Correct: C - Forget gate
Rationale:The forget gate f_t multiplies the previous cell state C_{t-1}, determining how much
old memory persists. The input gate controls new candidate information, the output gate
modulates the exposed hidden state, and 'reset gate' belongs to GRUs, not LSTMs.
Why the other answers are wrong:
A. The input gate scales the new candidate memory, not the retention of the old cell state.
B. The output gate controls how much of the cell state is exposed as the hidden state h_t.
D. Reset gates are components of GRU cells, not standard LSTM cells.
Reference: Hochreiter & Schmidhuber (1997). Long Short-Term Memory, Neural Computation; Colah's
LSTM blog.
Q4.
Which statement best characterizes the inductive bias of the self-attention mechanism
compared to a convolutional layer?
A. Self-attention is translation-equivariant by B. Self-attention has no built-in locality or
construction, while convolution is translation bias and must learn positional
permutation-invariant. information, whereas convolution
hard-codes local weight sharing.
C. Self-attention enforces a fixed receptive D. Self-attention requires fewer parameters
field per layer equal to its window size, like than convolution for the same sequence
convolution. length.
Correct: B - Self-attention has no built-in locality or translation bias and must learn
positional information, whereas convolution hard-codes local weight sharing.
Page 4
