CNN Architectures and Applications

Question 1

What is the primary problem with Fully Connected Networks (FCNs) regarding image processing?

Answer 1

Images are flattened into 1D vectors, losing spatial structure

This loss of spatial structure makes it difficult for FCNs to effectively understand visual data.

Question 2

How do Convolutional Neural Networks (CNNs) address the issues of FCNs?

Answer 2

Preserve spatial locality and handle images more intelligently

CNNs maintain the spatial relationships in images, improving accuracy.

Question 3

What is a Convolutional Neural Network (CNN)?

Answer 3

A deep learning model that works especially well with visual data like images

CNNs are specifically designed to process pixel data.

Question 4

What is the key idea behind how CNNs function?

Answer 4

They learn patterns in an image using filters

Filters help in detecting features such as edges and textures.

Question 5

What is the convolution operation in CNNs?

Answer 5

A small matrix called a kernel slides over the image, multiplying and summing to create a feature map

This operation extracts features from the input image.

Question 6

What does the term ‘stride’ refer to in the context of CNNs?

Answer 6

How far the filter moves each step

Stride affects the size of the output feature map.

Question 7

What are the two types of padding in CNNs?

Answer 7

‘valid’ (no padding) and ‘same’ (padding with zeros)

Padding is used to control the spatial dimensions of the output.

Question 8

What is the function of the convolutional layer in a CNN architecture?

Answer 8

Extracts feature maps

Convolutional layers are crucial for identifying patterns in images.

Question 9

What does the pooling layer do in a CNN?

Answer 9

Shrinks size while keeping important information

Pooling helps reduce the computational load and overfitting.

Question 10

What is the purpose of the activation layer in a CNN?

Answer 10

Adds non-linearity (e.g., ReLU)

Activation functions enable CNNs to learn complex patterns.

Question 11

What is the role of the fully connected layer in CNNs?

Answer 11

Final decision-making (e.g., classification)

Fully connected layers integrate features learned by previous layers.

Question 12

How do CNNs learn?

Answer 12

By adjusting filters to minimize prediction error

Learning occurs through a process called backpropagation.

Question 13

What is the basic process of CNN learning?

Answer 13

Input image goes through convolution + pooling, output passed to fully connected layers, loss function calculates error, backpropagation computes gradients, optimizer updates weights

This process iteratively improves the model’s predictions.

Question 14

What is an example of the input and output of a convolutional layer?

Answer 14

Input: 32×32 image, Output: 30×30×16

The output dimensions depend on the kernel size and stride.

Question 15

What are the total learnable parameters in a CNN example provided?

Answer 15

255,632

The number of parameters affects the model’s capacity to learn.

Question 16

What is max pooling in CNNs?

Answer 16

Picks the biggest number in a region

Max pooling helps retain the most significant features.

Question 17

What is average pooling in CNNs?

Answer 17

Takes the average of numbers in a region

Average pooling provides a smoother representation of features.

Question 18

What are the benefits of pooling layers in CNNs?

Answer 18

• Reduces size
• Controls overfitting
• Speeds up training

Pooling layers are essential for efficient training and model performance.

Question 19

What happens in the fully connected layer of a CNN?

Answer 19

All features combine to make a decision

This layer synthesizes information from all previous layers.

Question 20

What do convolutional layers learn using?

Answer 20

Filters

Filters are crucial for feature extraction.

Question 21

What is the purpose of pooling in CNNs?

Answer 21

Downsample while keeping key info

Pooling reduces dimensionality, making the model more efficient.

Question 22

What does activation introduce in CNNs?

Answer 22

Non-linear twist (usually ReLU)

Non-linear activations allow CNNs to model complex functions.

Question 23

What is the role of fully connected layers in CNNs?

Answer 23

Combine all for final output

Fully connected layers determine the final classification or regression output.

Question 24

What is the function of filters (kernels) in convolutional layers?

Answer 24

Scan over input image and create feature maps

Filters are essential for identifying patterns in the data.

Question 25

What is stride in the context of convolutional layers?

Answer 25

How far the filter moves ## Footnote Stride impacts the output size of the feature map.

Question 26

What does 'valid' padding mean?

Answer 26

No padding, smaller output ## Footnote Valid padding reduces the spatial dimensions of the feature map.

Question 27

What does 'same' padding do?

Answer 27

Keeps input/output sizes the same (zero-padding edges) ## Footnote Same padding allows for consistent output sizes.

Question 28

What is the typical structure of each convolutional layer?

Answer 28

Convolution and Activation (e.g., ReLU) ## Footnote This structure is fundamental for feature extraction.

Question 29

What is the most common pooling method used in CNNs?

Answer 29

MaxPooling2D ## Footnote Max pooling is widely used due to its effectiveness in feature retention.

Question 30

What is the pool size and stride for MaxPooling2D?

Answer 30

Pool size: (2,2), stride: (2,2) ## Footnote This configuration is standard for reducing dimensions.

Question 31

What do fully connected layers do with learned features?

Answer 31

Flatten them and feed into final classifier ## Footnote Flattening transforms the multi-dimensional output into a single vector.

Question 32

What is LeNet known for?

Answer 32

An early CNN architecture for digit recognition (MNIST) ## Footnote LeNet laid the groundwork for many modern CNN architectures.

Question 33

What is the key idea behind LeNet architecture?

Answer 33

Alternating Conv → ReLU → Pool ## Footnote This sequence is fundamental for effective feature extraction.

Question 34

What do feature maps in LeNet use to save memory?

Answer 34

Shared weights ## Footnote Sharing weights reduces the number of parameters in the model.

Question 35

What is the training configuration for the LeNet model?

Answer 35

20 epochs, Batch size: 128, Optimizer: Adam ## Footnote This configuration is typical for training CNNs.

Question 36

What are two regularization techniques mentioned for improving generalization?

Answer 36

* Spatial Dropout * Hyperparameter Tuning ## Footnote These techniques help prevent overfitting and improve model performance.

Question 37

What problem can arise during CNN training?

Answer 37

Vanishing Gradients or Exploding Gradients ## Footnote These issues can hinder the training process significantly.

Question 38

What does batch normalization do?

Answer 38

Makes training smoother, faster, and more stable ## Footnote Batch normalization helps in stabilizing the learning process.

Question 39

What is the formula for batch normalization?

Answer 39

Norm = (x - mean) / sqrt(var + ε) Out = γ * Norm + β ## Footnote This formula standardizes the input for each mini-batch.

Question 40

What are the benefits of using batch normalization?

Answer 40

* Reduces sensitivity to weight init * Slightly regularises model (less overfitting) * Allows higher learning rates ## Footnote Batch normalization enhances the overall performance of the model.

Question 41

When is it recommended to add batch normalization in Keras?

Answer 41

Before the activation ## Footnote Placing batch normalization before activation functions can lead to better results.

Question 42

What do convolutional layers do in a CNN?

Answer 42

Extract features ## Footnote Convolutional layers are vital for identifying key patterns in the input data.

Question 43

What is the purpose of pooling layers?

Answer 43

Shrink spatial size ## Footnote Pooling layers help reduce computational complexity.

Question 44

What is the function of fully connected layers?

Answer 44

Make decisions ## Footnote Fully connected layers synthesize information from previous layers to output final predictions.

Question 45

What do activation functions add to a neural network?

Answer 45

Non-linearity ## Footnote Non-linear activation functions are crucial for learning complex relationships.

Question 46

What does batch normalization do during training?

Answer 46

Stabilises training ## Footnote Batch normalization helps in maintaining consistent learning dynamics.

Question 47

What is transfer learning?

Answer 47

Reusing a pretrained model for your own task ## Footnote Transfer learning leverages existing knowledge to improve learning efficiency.

Question 48

What are the benefits of transfer learning?

Answer 48

* Saves time * Works well with small datasets * Leverages existing learned features ## Footnote Transfer learning is especially useful in scenarios with limited data.

Question 49

What is one strategy for transfer learning?

Answer 49

Freeze layers and only retrain the final classifier ## Footnote This strategy helps retain learned features while adapting to new tasks.

Question 50

What is data augmentation?

Answer 50

Pretend you have more data by slightly changing existing images ## Footnote Data augmentation helps improve model robustness.

Question 51

List examples of data augmentation techniques.

Answer 51

* Rotate * Flip * Zoom * Brightness tweak * Crop ## Footnote These techniques enhance the diversity of the training dataset.

Question 52

What does dropout do during training?

Answer 52

Randomly 'turn off' neurons ## Footnote Dropout helps prevent overfitting by ensuring that the model does not rely too heavily on any individual neuron.

Question 53

What problem does dropout help prevent?

Answer 53

* Overfitting * Reliance on specific paths in the network ## Footnote Dropout encourages a more generalized model.

Question 54

What is a common value for dropout?

Answer 54

Dropout(0.5) → 50% of neurons are dropped ## Footnote This value is often used to balance model complexity and performance.

Question 55

What is a characteristic of VGG architecture?

Answer 55

Uses only 3×3 convolutions ## Footnote VGG architecture is known for its simplicity and depth.

Question 56

What notable achievement is associated with AlexNet?

Answer 56

Won ImageNet 2012 ## Footnote AlexNet marked a significant advancement in deep learning for image classification.

Question 57

What is a key feature of ResNet architecture?

Answer 57

Uses skip connections (identity mappings) ## Footnote Skip connections help in training very deep networks by mitigating the vanishing gradient problem.

Question 58

What does ReLU stand for and what does it do?

Answer 58

ReLU: max(0, x); Fast and simple activation function ## Footnote ReLU is widely used due to its efficiency and effectiveness in learning.