Convolutional neural networks (CNNs) quiz

Question 1

What is the core operation that gives Convolutional Neural Networks their name, and what does it compute?

Accepted Answer

Sliding a filter over the input and computing weighted sums to form feature maps

Answer

Sorting pixel values from smallest to largest

Answer

Multiplying every pixel by a single global weight

Answer

Randomly shuffling pixels to remove noise

Question 2

Why are CNNs generally preferred over fully connected networks for image tasks?

Accepted Answer

They use weight sharing and local connectivity, reducing parameters

Answer

They never need any training data

Answer

They only work on black-and-white images

Answer

They eliminate the need for activation functions

Question 3

What is a feature map (activation map) in a CNN?

Accepted Answer

The output grid showing where a filter's pattern is detected

Answer

A list of the network's learning rate values

Answer

The raw RGB image before processing

Answer

The final classification probabilities

Question 4

What role does the pooling layer (e.g., max pooling) play in a CNN?

Accepted Answer

It downsamples feature maps, reducing size while keeping strong features

Answer

It increases the image resolution

Answer

It adds random noise for regularization

Answer

It converts the image to grayscale

Question 5

A 32x32 image is convolved with a 5x5 filter using stride 1 and no padding. What is the output spatial size?

Accepted Answer

28x28

Answer

30x30

Answer

32x32

Answer

27x27

Question 6

What is padding in a convolution, and what is 'same' padding?

Accepted Answer

Adding border pixels so output stays same size as input

Answer

Removing the border pixels of the image

Answer

Doubling the number of filters

Answer

Applying pooling twice in a row

Question 7

What does the stride of a convolution control, and what happens when you increase it?

Accepted Answer

Step size of the filter; larger stride gives smaller output

Answer

Number of filters in the layer

Answer

The learning rate of the optimizer

Answer

The number of training epochs

Question 8

Why is the ReLU activation function commonly used after convolution layers in CNNs?

Accepted Answer

It adds non-linearity, is cheap, and reduces vanishing gradients

Answer

It guarantees the output sums to one

Answer

It removes the need for convolution

Answer

It always outputs negative values

Question 9

In CNNs, what does 'translation invariance' mean and which component helps achieve it?

Accepted Answer

Recognizing features regardless of position; aided by weight sharing and pooling

Answer

Recognizing only centered objects; aided by padding

Answer

Translating images into another language

Answer

Rotating the image before classification

Question 10

How does a convolutional filter handle a color (RGB) image with 3 channels?

Accepted Answer

The filter has depth equal to input channels and sums across them

Answer

It processes each channel as a separate output, never combining

Answer

It ignores two channels and uses only red

Answer

It converts RGB to a single gray value first

Question 11

What is the difference between the number of filters and the size of a filter in a conv layer?

Accepted Answer

Size is each kernel's extent; number of filters sets output depth

Answer

They are the same thing

Answer

Number of filters sets the image width

Answer

Size sets how many epochs to train

Question 12

How many learnable parameters does a convolutional layer with 16 filters of size 3x3 applied to a 3-channel input have (including biases)?

Accepted Answer

448

Answer

432

Answer

144

Answer

160

Question 13

What is the receptive field of a neuron in a CNN, and how does it change with depth?

Accepted Answer

Input region affecting a neuron; it grows with depth

Answer

The number of filters in a layer; it stays constant

Answer

The learning rate region; it shrinks with depth

Answer

The output image size; it never changes

Question 14

What is the typical role of the fully connected (dense) layers at the end of a classification CNN?

Accepted Answer

They combine extracted features to produce final class scores

Answer

They perform the convolution operation

Answer

They downsample the feature maps

Answer

They add padding to the input

Question 15

Why is the softmax function used in the output layer of a multi-class CNN classifier?

Accepted Answer

It turns scores into probabilities that sum to 1 across classes

Answer

It removes negative pixel values

Answer

It downsamples feature maps

Answer

It increases the receptive field

Question 16

What hierarchy of features do early versus deep layers in a CNN typically learn?

Accepted Answer

Early: edges/textures; deep: whole objects/concepts

Answer

Early: whole objects; deep: single pixels

Answer

All layers learn the same features

Answer

Deep layers only learn colors

Question 17

What is data augmentation and why is it useful when training CNNs on images?

Accepted Answer

Creating transformed copies of images to improve generalization

Answer

Deleting half the training images

Answer

Reducing the number of filters

Answer

Increasing the learning rate over time

Question 18

What is transfer learning in the context of CNNs?

Accepted Answer

Reusing a pretrained CNN and adapting it to a new task

Answer

Training a CNN entirely from random weights every time

Answer

Transferring images between two computers

Answer

Converting a CNN into a decision tree

Question 19

What problem does batch normalization help solve in deep CNNs?

Accepted Answer

It stabilizes layer input distributions, speeding up training

Answer

It increases image resolution

Answer

It removes the need for filters

Answer

It performs the final classification

Question 20

How does dropout act as a regularizer in CNNs, and where is it commonly placed?

Accepted Answer

Randomly zeroing activations during training to reduce overfitting

Answer

Permanently deleting random filters from the model

Answer

Increasing all activations by a fixed amount

Answer

Adding extra pixels to the image

Question 21

What is the vanishing gradient problem and how do residual connections (ResNets) help?

Accepted Answer

Gradients shrink in deep nets; skip connections let gradients flow

Answer

Gradients explode; pooling fixes it

Answer

Filters become too large; padding fixes it

Answer

The image becomes blurry; softmax fixes it

Question 22

What is max pooling versus average pooling, and how do their outputs differ?

Accepted Answer

Max takes the largest value; average takes the mean of the window

Answer

Both always return zero

Answer

Max takes the mean; average takes the largest value

Answer

They are identical operations

Question 23

What is the purpose of flattening between the convolutional part and the dense part of a CNN?

Accepted Answer

It converts 3D feature maps into a 1D vector for dense layers

Answer

It performs convolution on the output

Answer

It removes all but one channel

Answer

It increases the spatial resolution

Question 24

Why might a 3x3 convolution stacked twice be preferred over a single 5x5 convolution?

Accepted Answer

Same receptive field, fewer parameters, and extra non-linearity

Answer

It is slower but uses more parameters

Answer

It removes the need for activation functions

Answer

It can only process grayscale images

Question 25

What loss function is typically used to train a multi-class CNN image classifier?

Accepted Answer

Categorical cross-entropy

Answer

Mean squared error of pixel values

Answer

Filter overlap distance

Answer

Pooling variance loss

Question 26

If a max pooling layer with a 2x2 window and stride 2 is applied to a 28x28 feature map, what is the output size?

Accepted Answer

14x14

Answer

28x28

Answer

7x7

Answer

26x26

Question 27

What does a global average pooling (GAP) layer do, and why is it used before the output?

Accepted Answer

Averages each feature map to one value, reducing parameters

Answer

Doubles the spatial resolution

Answer

Adds noise to each channel

Answer

Performs the convolution operation

Question 28

In a CNN, what is meant by the 'depth' of a convolutional layer's output?

Accepted Answer

The number of filters, equal to the number of output feature maps

Answer

The height of the input image

Answer

The number of training epochs

Answer

The size of each filter

Question 29

Write a small PyTorch snippet defining a convolutional layer that takes a 3-channel image and outputs 16 feature maps using 3x3 filters with same padding.

Accepted Answer

nn.Conv2d(3, 16, kernel_size=3, padding=1)

Answer

nn.Linear(3, 16)

Answer

nn.MaxPool2d(3, 16)

Answer

nn.Conv2d(16, 3, kernel_size=0)

Question 30

Why does a CNN typically increase the number of filters while decreasing spatial size as you go deeper?

Accepted Answer

To trade spatial detail for richer high-level feature representation

Answer

Because filters must always equal image width

Answer

To make the image larger

Answer

Because deeper layers cannot use filters

Test yourself