CS 7641 Final Exam Practice | Questions
and Answers | 2026 Update | 100% Correct -
GT.
Practice Question Bank & Answer Rationale
THIS PACK CONTAINS
148 exam-style multiple-choice questions. Each question includes the correct answer and a complete
rationale for focused revision.
COVERAGE
This exam covers supervised learning, unsupervised learning, reinforcement learning, and learning theory at
the graduate level. It emphasizes conceptual understanding, algorithmic reasoning, and practical application
of machine learning techniques.
STUDY GUIDE
Recommended duration: 3 hours. Passing target: 70%.
HIGHP - Page 1 of 48
,1. In a decision tree, which attribute selection measure is most sensitive to the number of
attribute values?
[ ] A. Information Gain
[ ] B. Gain Ratio
[ ] C. Gini Index
[ ] D. Classification Error
CORRECT: A. Information Gain
Information Gain is biased toward attributes with many values because it tends to overfit by creating many
small partitions. Gain Ratio was introduced to correct this bias, while Gini Index and Classification Error are
less sensitive.
2. In k-means clustering, which of the following is true about the objective function?
[ ] A. It converges to the global minimum for any initialization.
[ ] B. It monotonically decreases with each iteration until convergence.
[ ] C. It is guaranteed to find the optimal number of clusters.
[ ] D. It requires the clusters to be of equal size.
CORRECT: B. It monotonically decreases with each iteration until convergence.
The k-means objective (sum of squared distances to centroids) is non-increasing across iterations because
each step reduces or maintains the objective. However, it may converge to a local minimum depending on
initialization.
3. In Q-learning, what is the purpose of the learning rate parameter ?
[ ] A. It determines the discount factor for future rewards.
[ ] B. It controls the trade-off between exploration and exploitation.
[ ] C. It weights the importance of new information versus existing Q-values.
[ ] D. It defines the maximum number of iterations.
CORRECT: C. It weights the importance of new information versus existing Q-values.
The learning rate (0 < 1) determines how much the newly acquired information overrides the old Q-value. A
value of 0 means the agent learns nothing, while 1 means it only considers the most recent information.
4. Which of the following statements about the VC dimension is correct?
[ ] A. It measures the number of training examples a model can memorize.
[ ] B. It is the maximum number of points that can be shattered by a hypothesis class.
[ ] C. It is always equal to the number of parameters in the model.
[ ] D. It determines the exact generalization error of a model.
HIGHP - Page 2 of 48
, CORRECT: B. It is the maximum number of points that can be shattered by a hypothesis class.
The VC dimension is defined as the largest number of points that can be shattered (i.e., correctly classified in
all possible labelings) by a hypothesis class. It provides a measure of capacity and informs generalization
bounds.
5. In a neural network, which of the following is a consequence of using a very high
learning rate?
[ ] A. The model may converge too slowly.
[ ] B. The model may oscillate or diverge.
[ ] C. The model will always overfit.
[ ] D. The model will underfit.
CORRECT: B. The model may oscillate or diverge.
A high learning rate can cause the optimization to overshoot the minimum, leading to oscillations or
divergence. It does not directly cause overfitting or underfitting, and slow convergence is typically due to a
low learning rate.
6. Which of the following is a key difference between batch gradient descent and
stochastic gradient descent?
[ ] A. Batch gradient descent uses a subset of data for each update.
[ ] B. Stochastic gradient descent updates parameters using a single training example.
[ ] C. Batch gradient descent is always faster to converge.
[ ] D. Stochastic gradient descent cannot handle large datasets.
CORRECT: B. Stochastic gradient descent updates parameters using a single training example.
Stochastic gradient descent (SGD) updates parameters using a single training example per iteration, while
batch gradient descent uses the entire dataset. SGD can be faster per iteration and can handle large
datasets by processing one example at a time.
7. In the context of reinforcement learning, what does the Bellman equation represent?
[ ] A. The relationship between the value of a state and the values of its successor states.
[ ] B. The optimal policy for a given environment.
[ ] C. The exploration-exploitation trade-off.
[ ] D. The convergence criterion for Q-learning.
CORRECT: A. The relationship between the value of a state and the values of its successor states.
The Bellman equation expresses the value of a state in terms of the immediate reward and the discounted
value of successor states. It is fundamental to dynamic programming and reinforcement learning algorithms.
HIGHP - Page 3 of 48
, 8. Which of the following is an advantage of using ensemble methods like Random
Forests over a single decision tree?
[ ] A. They are easier to interpret.
[ ] B. They reduce variance without increasing bias.
[ ] C. They require less computational resources.
[ ] D. They can only handle binary classification.
CORRECT: B. They reduce variance without increasing bias.
Random Forests combine multiple decorrelated trees to reduce variance while maintaining low bias, often
improving generalization. They are typically less interpretable and more computationally intensive than a
single tree.
9. In principal component analysis (PCA), what is the primary goal?
[ ] A. To maximize the variance of the projected data.
[ ] B. To minimize the reconstruction error.
[ ] C. To find the most discriminative directions.
[ ] D. To reduce the number of features while preserving class separability.
CORRECT: A. To maximize the variance of the projected data.
PCA seeks to project data onto directions (principal components) that maximize variance, which is equivalent
to minimizing reconstruction error. It is unsupervised and does not consider class labels, unlike LDA.
10. Which of the following statements about the bias-variance tradeoff is correct?
[ ] A. Increasing model complexity typically increases bias and decreases variance.
[ ] B. Increasing model complexity typically decreases bias and increases variance.
[ ] C. The bias-variance tradeoff only applies to regression, not classification.
[ ] D. Regularization increases both bias and variance.
CORRECT: B. Increasing model complexity typically decreases bias and increases variance.
As model complexity increases, the model can fit the training data more closely, reducing bias but increasing
variance (overfitting). Regularization typically increases bias but decreases variance.
11. In a decision tree, which splitting criterion is most appropriate when the target
attribute is continuous and the tree is used for regression?
[ ] A. Information gain
[ ] B. Gain ratio
[ ] C. Variance reduction
[ ] D. Gini index
HIGHP - Page 4 of 48
and Answers | 2026 Update | 100% Correct -
GT.
Practice Question Bank & Answer Rationale
THIS PACK CONTAINS
148 exam-style multiple-choice questions. Each question includes the correct answer and a complete
rationale for focused revision.
COVERAGE
This exam covers supervised learning, unsupervised learning, reinforcement learning, and learning theory at
the graduate level. It emphasizes conceptual understanding, algorithmic reasoning, and practical application
of machine learning techniques.
STUDY GUIDE
Recommended duration: 3 hours. Passing target: 70%.
HIGHP - Page 1 of 48
,1. In a decision tree, which attribute selection measure is most sensitive to the number of
attribute values?
[ ] A. Information Gain
[ ] B. Gain Ratio
[ ] C. Gini Index
[ ] D. Classification Error
CORRECT: A. Information Gain
Information Gain is biased toward attributes with many values because it tends to overfit by creating many
small partitions. Gain Ratio was introduced to correct this bias, while Gini Index and Classification Error are
less sensitive.
2. In k-means clustering, which of the following is true about the objective function?
[ ] A. It converges to the global minimum for any initialization.
[ ] B. It monotonically decreases with each iteration until convergence.
[ ] C. It is guaranteed to find the optimal number of clusters.
[ ] D. It requires the clusters to be of equal size.
CORRECT: B. It monotonically decreases with each iteration until convergence.
The k-means objective (sum of squared distances to centroids) is non-increasing across iterations because
each step reduces or maintains the objective. However, it may converge to a local minimum depending on
initialization.
3. In Q-learning, what is the purpose of the learning rate parameter ?
[ ] A. It determines the discount factor for future rewards.
[ ] B. It controls the trade-off between exploration and exploitation.
[ ] C. It weights the importance of new information versus existing Q-values.
[ ] D. It defines the maximum number of iterations.
CORRECT: C. It weights the importance of new information versus existing Q-values.
The learning rate (0 < 1) determines how much the newly acquired information overrides the old Q-value. A
value of 0 means the agent learns nothing, while 1 means it only considers the most recent information.
4. Which of the following statements about the VC dimension is correct?
[ ] A. It measures the number of training examples a model can memorize.
[ ] B. It is the maximum number of points that can be shattered by a hypothesis class.
[ ] C. It is always equal to the number of parameters in the model.
[ ] D. It determines the exact generalization error of a model.
HIGHP - Page 2 of 48
, CORRECT: B. It is the maximum number of points that can be shattered by a hypothesis class.
The VC dimension is defined as the largest number of points that can be shattered (i.e., correctly classified in
all possible labelings) by a hypothesis class. It provides a measure of capacity and informs generalization
bounds.
5. In a neural network, which of the following is a consequence of using a very high
learning rate?
[ ] A. The model may converge too slowly.
[ ] B. The model may oscillate or diverge.
[ ] C. The model will always overfit.
[ ] D. The model will underfit.
CORRECT: B. The model may oscillate or diverge.
A high learning rate can cause the optimization to overshoot the minimum, leading to oscillations or
divergence. It does not directly cause overfitting or underfitting, and slow convergence is typically due to a
low learning rate.
6. Which of the following is a key difference between batch gradient descent and
stochastic gradient descent?
[ ] A. Batch gradient descent uses a subset of data for each update.
[ ] B. Stochastic gradient descent updates parameters using a single training example.
[ ] C. Batch gradient descent is always faster to converge.
[ ] D. Stochastic gradient descent cannot handle large datasets.
CORRECT: B. Stochastic gradient descent updates parameters using a single training example.
Stochastic gradient descent (SGD) updates parameters using a single training example per iteration, while
batch gradient descent uses the entire dataset. SGD can be faster per iteration and can handle large
datasets by processing one example at a time.
7. In the context of reinforcement learning, what does the Bellman equation represent?
[ ] A. The relationship between the value of a state and the values of its successor states.
[ ] B. The optimal policy for a given environment.
[ ] C. The exploration-exploitation trade-off.
[ ] D. The convergence criterion for Q-learning.
CORRECT: A. The relationship between the value of a state and the values of its successor states.
The Bellman equation expresses the value of a state in terms of the immediate reward and the discounted
value of successor states. It is fundamental to dynamic programming and reinforcement learning algorithms.
HIGHP - Page 3 of 48
, 8. Which of the following is an advantage of using ensemble methods like Random
Forests over a single decision tree?
[ ] A. They are easier to interpret.
[ ] B. They reduce variance without increasing bias.
[ ] C. They require less computational resources.
[ ] D. They can only handle binary classification.
CORRECT: B. They reduce variance without increasing bias.
Random Forests combine multiple decorrelated trees to reduce variance while maintaining low bias, often
improving generalization. They are typically less interpretable and more computationally intensive than a
single tree.
9. In principal component analysis (PCA), what is the primary goal?
[ ] A. To maximize the variance of the projected data.
[ ] B. To minimize the reconstruction error.
[ ] C. To find the most discriminative directions.
[ ] D. To reduce the number of features while preserving class separability.
CORRECT: A. To maximize the variance of the projected data.
PCA seeks to project data onto directions (principal components) that maximize variance, which is equivalent
to minimizing reconstruction error. It is unsupervised and does not consider class labels, unlike LDA.
10. Which of the following statements about the bias-variance tradeoff is correct?
[ ] A. Increasing model complexity typically increases bias and decreases variance.
[ ] B. Increasing model complexity typically decreases bias and increases variance.
[ ] C. The bias-variance tradeoff only applies to regression, not classification.
[ ] D. Regularization increases both bias and variance.
CORRECT: B. Increasing model complexity typically decreases bias and increases variance.
As model complexity increases, the model can fit the training data more closely, reducing bias but increasing
variance (overfitting). Regularization typically increases bias but decreases variance.
11. In a decision tree, which splitting criterion is most appropriate when the target
attribute is continuous and the tree is used for regression?
[ ] A. Information gain
[ ] B. Gain ratio
[ ] C. Variance reduction
[ ] D. Gini index
HIGHP - Page 4 of 48
