Why is generalization more important than training accuracy in neural networks?

•  Generalization proves convergence.

• Training accuracy ensures bias reduction.

•  It prevents vanishing gradients.

• It reflects the ability to predict unseen data, the true goal.

A multilayer network without nonlinearities collapses into a ______ model.

When using ReLU in hidden layers instead of sigmoid, which benefit typically emerges?

• 1. Guarantees linear separability in the input space.
     
     

• Prevents exploding gradients.

• Ensures all neurons remain active.

• Reduces vanishing gradient problems

Which of the following statements about gradient descent is correct?

• It always finds the global minimum.
  

•  It updates weights by moving against the gradient of the loss function.

• It requires linear separability.

• It is identical to the perceptron update rule.

Why is nonlinearity essential in deep neural networks?

• It guarantees zero error.
  

•  It reduces training time.

• It allows the composition of layers to model complex, non-linear boundaries.

• It simplifies the optimization problem.

Backpropagation uses the ______ rule to propagate gradients backward through layers

In a neural network, forward propagation refers to:

• Feeding inputs through the network to generate predictions

• Updating weights using gradient descent

• Reversing gradients to find errors

• Adjusting biases to prevent saturation

Which activation function is most commonly used in modern deep networks for hidden layers?

• Sigmoid

• Tanh

• ReLU

• Sign function

A perceptron without a bias term may struggle because:

• It cannot represent decision boundaries that don't pass through the origin

• It cannot compute loss correctly

• It requires too many hidden layers

• It cannot apply nonlinear activations

Which statement about cross-entropy loss is correct?

• It measures the squared difference between predicted and actual values

• It is used to calculate the probability of linear separability

• It penalizes wrong predictions by increasing loss logarithmically

• It works only with ReLU activations