Question 1 of 25

What does a perceptron compute before applying an activation?

Question 2 of 25

Why can a single perceptron not learn XOR?

Question 3 of 25

Why do we need activation functions in hidden layers?

Question 4 of 25

What is a key advantage of ReLU over sigmoid in hidden layers?

Question 5 of 25

Which activation is most common for binary output probability?

Question 6 of 25

Forward propagation means:

Question 7 of 25

What does backpropagation actually compute?

Question 8 of 25

What is the vanishing gradient problem?

Question 9 of 25

For multi-class digit classification (0–9), which loss is standard?

Question 10 of 25

Loss is high but not decreasing. What is a sensible first check?

Question 11 of 25

A network has input 784, hidden 128, output 10. How many weight connections (approx.) from input to hidden?

Question 12 of 25

Sigmoid output near 0 or 1 with saturated hidden sigmoids contributed to vanishing gradients because:

Question 13 of 25

After forward pass and loss, weights update using:

Question 14 of 25

Stacking multiple neurons in a hidden layer fixes XOR because:

Question 15 of 25

Why use softmax + cross-entropy together on the output layer for MNIST?

Question 16 of 25

A perceptron without activation on the output is limited because:

Question 17 of 25

Tanh activation outputs are in range:

Question 18 of 25

In a 3-layer network (input → hidden → output), forward propagation computes:

Question 19 of 25

Backpropagation is efficient because it:

Question 20 of 25

Why is ReLU often preferred over sigmoid in hidden layers?

Question 21 of 25

MSE loss is commonly used for:

Question 22 of 25

If a network has no activation between linear layers, stacking many layers is equivalent to:

Question 23 of 25

Learning rate too high during neural net training often causes:

Question 24 of 25

The perceptron computes weighted sum + bias, then applies:

Question 25 of 25

For binary classification (spam vs ham), a common output activation + loss pair is: