Deep Learning Specialization Final Exam 2026: University &
Professional Practice Test — Neural Network Foundations,
Optimization, Regularization, CNNs, Sequence Models, Transformers
& Model Deployment — 190 Multiple-Choice Questions with Answers
& Detailed Rationales
Section 1: Neural Network Foundations and Logistic Regression (Q1–Q25)
Question 1
In logistic regression for binary classification, what does the output of the sigmoid function
represent?
A. The unnormalized log-odds of the positive class.
B. The estimated probability that the input belongs to the positive class.
C. The margin between the decision boundary and the nearest point.
D. The squared error between prediction and target.
Correct Answer: B
Explanation: Sigmoid maps real-valued logits to (0,1), interpreted as P(y=1|x). It is not a
margin or error.
Question 2
A neural network with no hidden layers and a sigmoid output is equivalent to which model?
A. A linear regression model.
B. A logistic regression model.
C. A support vector machine.
D. A decision tree.
,Correct Answer: B
Explanation: A single sigmoid output unit with learned weights computes exactly the logistic
regression hypothesis.
Question 3
Why is the ReLU activation often preferred over sigmoid in deep hidden layers?
A. It bounds activations strictly between -1 and 1.
B. It reduces the vanishing gradient problem for positive inputs.
C. It guarantees zero mean activations.
D. It makes the network convex.
Correct Answer: B
Explanation: ReLU has derivative 1 for positive inputs, mitigating vanishing gradients. It does
not bound outputs or ensure zero mean.
Question 4
In a feedforward network, what is the role of the bias term in each neuron?
A. It normalizes the input features.
B. It shifts the activation function, allowing flexible decision boundaries.
C. It prevents overfitting by adding noise.
D. It ensures the output is always positive.
Correct Answer: B
,Explanation: Bias shifts the pre-activation, enabling the neuron to fire even when weighted
inputs are zero.
Question 5
Which quantity does the cross-entropy loss penalize most strongly?
A. Predictions that are correct and confident.
B. Predictions that are wrong and confident.
C. Predictions that are correct but uncertain.
D. Predictions that are uniformly random.
Correct Answer: B
Explanation: Cross-entropy is -log(p_true); when p_true is small, loss becomes very large,
heavily penalizing confident errors.
Question 6
A student computes gradients for a 3-layer network but forgets to cache intermediate
activations. What is the most likely consequence?
A. Backpropagation cannot compute gradients correctly.
B. The network becomes a linear model.
C. The learning rate must be doubled.
D. The loss becomes convex.
Correct Answer: A
, Explanation: Backpropagation relies on cached forward activations to compute chain-rule
derivatives; omitting them breaks gradient computation.
Question 7
What is the primary purpose of vectorizing logistic regression over examples?
A. To reduce the number of parameters.
B. To exploit parallel hardware and avoid explicit loops.
C. To eliminate the need for a bias term.
D. To make the loss non-convex.
Correct Answer: B
Explanation: Vectorization replaces per-example loops with matrix operations, enabling
efficient GPU/CPU parallelism.
Question 8
In a binary classification network, the output layer uses sigmoid. Which loss is most
appropriate?
A. Mean squared error with linear output.
B. Binary cross-entropy.
C. Categorical cross-entropy with softmax.
D. Hinge loss with margin.
Correct Answer: B
Explanation: Binary cross-entropy matches a Bernoulli likelihood and pairs naturally with
sigmoid outputs.
Professional Practice Test — Neural Network Foundations,
Optimization, Regularization, CNNs, Sequence Models, Transformers
& Model Deployment — 190 Multiple-Choice Questions with Answers
& Detailed Rationales
Section 1: Neural Network Foundations and Logistic Regression (Q1–Q25)
Question 1
In logistic regression for binary classification, what does the output of the sigmoid function
represent?
A. The unnormalized log-odds of the positive class.
B. The estimated probability that the input belongs to the positive class.
C. The margin between the decision boundary and the nearest point.
D. The squared error between prediction and target.
Correct Answer: B
Explanation: Sigmoid maps real-valued logits to (0,1), interpreted as P(y=1|x). It is not a
margin or error.
Question 2
A neural network with no hidden layers and a sigmoid output is equivalent to which model?
A. A linear regression model.
B. A logistic regression model.
C. A support vector machine.
D. A decision tree.
,Correct Answer: B
Explanation: A single sigmoid output unit with learned weights computes exactly the logistic
regression hypothesis.
Question 3
Why is the ReLU activation often preferred over sigmoid in deep hidden layers?
A. It bounds activations strictly between -1 and 1.
B. It reduces the vanishing gradient problem for positive inputs.
C. It guarantees zero mean activations.
D. It makes the network convex.
Correct Answer: B
Explanation: ReLU has derivative 1 for positive inputs, mitigating vanishing gradients. It does
not bound outputs or ensure zero mean.
Question 4
In a feedforward network, what is the role of the bias term in each neuron?
A. It normalizes the input features.
B. It shifts the activation function, allowing flexible decision boundaries.
C. It prevents overfitting by adding noise.
D. It ensures the output is always positive.
Correct Answer: B
,Explanation: Bias shifts the pre-activation, enabling the neuron to fire even when weighted
inputs are zero.
Question 5
Which quantity does the cross-entropy loss penalize most strongly?
A. Predictions that are correct and confident.
B. Predictions that are wrong and confident.
C. Predictions that are correct but uncertain.
D. Predictions that are uniformly random.
Correct Answer: B
Explanation: Cross-entropy is -log(p_true); when p_true is small, loss becomes very large,
heavily penalizing confident errors.
Question 6
A student computes gradients for a 3-layer network but forgets to cache intermediate
activations. What is the most likely consequence?
A. Backpropagation cannot compute gradients correctly.
B. The network becomes a linear model.
C. The learning rate must be doubled.
D. The loss becomes convex.
Correct Answer: A
, Explanation: Backpropagation relies on cached forward activations to compute chain-rule
derivatives; omitting them breaks gradient computation.
Question 7
What is the primary purpose of vectorizing logistic regression over examples?
A. To reduce the number of parameters.
B. To exploit parallel hardware and avoid explicit loops.
C. To eliminate the need for a bias term.
D. To make the loss non-convex.
Correct Answer: B
Explanation: Vectorization replaces per-example loops with matrix operations, enabling
efficient GPU/CPU parallelism.
Question 8
In a binary classification network, the output layer uses sigmoid. Which loss is most
appropriate?
A. Mean squared error with linear output.
B. Binary cross-entropy.
C. Categorical cross-entropy with softmax.
D. Hinge loss with margin.
Correct Answer: B
Explanation: Binary cross-entropy matches a Bernoulli likelihood and pairs naturally with
sigmoid outputs.
