ANN Lecture 5 - Convolutional Neural Networks

Question 1

What is wrong with using only fully connected layers for image classification?

Answer 1

Local patterns are not used, because the input images get flattened into one long array.

Question 2

Problem of Template Matching

Answer 2

Only matches template and image part if identical. If there is one pixel of noise it is no match.

Question 3

Cross Corelation

Answer 3

Slide a kernel over a signal and compute the dot product.

Question 4

Convolutional Neural Network - Inputshape

Answer 4

Batchsize, height, width, channels

Question 5

Convolutional Neural Network - Convolutional Layer

Answer 5

• The kernels that slide over the image get learned. That means the kernel values are the learnable weights.
• For each kernel we get one feature map
• All the neurons in one feature map share the same weights
• each neuron will have different activations as they “see” different parts of the input

Question 6

CNN - Convolution Layer Parameters

Answer 6

• The number of different kernes (which define the number of feature maos in the output)
• The shape of the kernes: The shape will be (kernel size, lernel size, channels of input to this layer)
• The padding: valid or same(Feature map has same size as the input)
• The stride size: how far do you slide the kernel per step

Question 7

Pooling

Answer 7

Pooling helps reducing the size of the feature maps and generalizing by loosing the exact position of the information.
Max Pooling: Taking the maximum value of the kernel
Average Pooling: Taking the average over all values of the kernel

Question 8

Read-Out-Network

Answer 8

• After multiple convolution and pooling layers the size of the feature maps are reduced a lot
• The feature maps get flattened to one vector
• One or more fully connected layers are applied