Module 14 2022

01/04/2022

In theory, early filters look for basic features (edges, corners etc) and later filters look at the maps of these features to identify groups of features.

These filters do not need to be designed manually! The same backpropagation process that adjusts weights in a classic neural network will also adjust filters in a convolutional neural network, so the network “learns” the features inherent in the training data.

Image © Nvidia

The Organisation for Professionals in Regulatory Affairs

13

Example: Reading road signs

Through convolution and the use of multiple layers with multiple filters, a numerical function is created which describes the features of the image. This is then used as the input for a neural network, at dramatically lower computational cost than using each pixel as an input.

Image © Nvidia

The Organisation for Professionals in Regulatory Affairs

14

7