Module 14 2022

01/04/2022

The problem

These sorts of neural networks alone are not practical when working with many datasets.

The number of weight and bias variables rapidly becomes unmanageable, for example a 1MP colour image would require 3 million discrete inputs in the first layer.

Video © 3blue1brown.com

The Organisation for Professionals in Regulatory Affairs

9

Image Analysis Using Convolution

Convolutional neural networks work by applying a small (in this example 3x3 pixels) filter to every point in an image, and collecting the output. These filters detect properties of the input image, in this case vertical edges. Multiple different filters are applied to the same image, creating maps of different features.

Image © datahacker.rs

The Organisation for Professionals in Regulatory Affairs

10

5