CNN Batch Normalization and Dropout 2

Question 1

What does a dropout layer do during training?

Accepted Answer

Randomly sets a fraction of neuron activations to zero

Answer

Randomly shuffles the order of filters in a convolutional layer

Answer

Reduces the learning rate for randomly selected weights

Answer

Clips gradient values above a threshold to zero

Question 2

What happens to dropout during model inference/testing?

Accepted Answer

All neurons are active and outputs are scaled by the keep probability

Answer

Dropout is applied with half the training dropout rate

Answer

Dropout continues but only in the final fully connected layer

Answer

A new random mask is sampled and applied once at inference time

Question 3

What is 'inverted dropout' and why is it preferred in practice?

Accepted Answer

Scaling kept activations by 1/(1-p) during training so no scaling is needed at test time

Answer

Applying dropout only to the input layer and inverting the mask at deeper layers

Answer

Using a dropout rate that increases with network depth

Answer

Randomly scaling activations up instead of zeroing them down

Question 4

How does dropout act as a form of regularization?

Accepted Answer

By forcing the network to learn robust features not dependent on specific co-adaptations

Answer

By penalizing large weight values via an L2 term in the loss function

Answer

By reducing the number of parameters in the network permanently

Answer

By increasing the effective learning rate to escape sharp minima

Question 5

What is spatial dropout (also called 2D dropout) and when is it used?

Accepted Answer

Dropping entire feature maps (channels) rather than individual activations in CNNs

Answer

Applying dropout only in the spatial height/width dimensions but keeping all channels

Answer

Dropping activations in a structured grid pattern across the feature map

Answer

Randomly masking contiguous rectangular regions of the input image

Question 6

What dropout rate is commonly used as a starting default in fully connected layers of CNNs?

Accepted Answer

0.5 (drop 50% of neurons)

Answer

0.1 (drop 10% of neurons)

Answer

0.9 (drop 90% of neurons)

Answer

0.01 (drop 1% of neurons)

Question 7

Which interpretation explains why dropout works from an ensemble learning perspective?

Accepted Answer

Each forward pass trains a different thinned network, and inference averages these networks

Answer

Dropout creates multiple copies of the network that are trained on different subsets of data

Answer

Dropped neurons act as negative examples that improve contrastive learning

Answer

Dropout forces the optimizer to use different learning rates for each layer