Deep Learning 1: Foundations of Convolutional Neural Networks

Convolutional neural networks have three types of layers: Convolutional layer, Pooling layer, Fully connected layer. Usually, a convolutional layer and a pooling layer are called the same layer since a pooling layer doesn't have parameters remained to be fitted.

 

1. Convolutional layer 

Parameters: weights Deep Learning 1: Foundations of Convolutional Neural Networks and bias Deep Learning 1: Foundations of Convolutional Neural Networks

Hyperparameters: padding Deep Learning 1: Foundations of Convolutional Neural Networks, stride Deep Learning 1: Foundations of Convolutional Neural Networks, filter size Deep Learning 1: Foundations of Convolutional Neural Networks, filter number Deep Learning 1: Foundations of Convolutional Neural Networks

  • Padding and stride 

Convolution without padding has two problems:

  1. shrink output 
  2. throw away information from the edge of the image

Stride decreases the displacement of images.

Deep Learning 1: Foundations of Convolutional Neural Networks

Deep Learning 1: Foundations of Convolutional Neural Networks

Deep Learning 1: Foundations of Convolutional Neural Networks

  • Convolutionover volume

An RBG image using a filter becomes an image with one channel.

Deep Learning 1: Foundations of Convolutional Neural Networks

  • Example: one convolutional layer 

Deep Learning 1: Foundations of Convolutional Neural Networks

 

2. Pooling layer 

Pooling layer can be considered as a max/average filter.

Deep Learning 1: Foundations of Convolutional Neural Networks

 

3. Fully connected layer 

Deep Learning 1: Foundations of Convolutional Neural Networks

4. LeNet-5

Deep Learning 1: Foundations of Convolutional Neural Networks