Overview

• CNNs are uncommonly good at finding features (in an image).
• They have a hierarchal way of seeing things; inspired by examining the way our own visual cortex works.
• A convolutional layer is in contrast to a fully connected layer in that each neuron is only connected to a few nearby local neurons in the previous layer.

Mathematical Background

• The kernel/filter is just a matrix of weights, like the example below:
• The input image is also pre-processed into a matrix of weights, although larger in dimensionality than the kernel.

Process of Operation

• Convolutional layer
• Nonlinarity layer
• Pooling layer

Networks that contain only the above three layers are known as 'Fully Convolutional Networks'

• Fully connected layer

• Once trained, new images are classified by inferring their similarities to the trained model.

Training

• CNNs will receive a big, labelled, data set with many images of the type of object(s) they aim to recognise.
• The outputs of the network are compared to what is expected (possible due to the fact that the training data is labelled).
• The cost function that is used in this comparison, depends on the task:
• Once loss is measured, the back propagation algorithm will calculate the partial derivatives with regards to the neural network parameters. Then a regular gradient descent method will be used.