Week4_S2

Max pooling with a 2x2 window and stride 2 reduces the output size by:

•  Half in each dimension

• Quarter in each dimension

• No change in size

• One-third in each dimension

The output produced by applying a filter to an image is called a __________ map.

To prevent overfitting, pooling layers reduce the number of __________ in the network.

What is the main role of pooling layers in a CNN?

• Decrease model complexity
  

• Downsample feature maps

• Add more parameters to the network

•  Improve filter weights directly

Which of the following combinations will most likely lead to overfitting in a CNN?

• Small filters + pooling + dropout

• Large filters + small dataset + no regularization

• Many small filters + large dataset

• High learning rate + regularization

The Flatten() layer is used to:

• Normalize pixel values

•  Reduce the number of features

•  Convert feature maps into a vector
  

• Increase batch size

Increasing the number of convolutional filters primarily:

•  Reduces model complexity

• Replaces pooling layers

• Reduces computation time

•  Increases model’s feature extraction capacity

Zero-padding is often used to:

• Reduce overfitting

•  Increase the input image dimension

•  Randomly remove neurons

• Preserve feature map dimensions

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

• Sigmoid

• ReLU

• Tanh

• Softmax

__________ pooling keeps the maximum value from each region of the input.