A simple one-layer perceptron, which can tell if a hand written image is a 3 or not.
Goal: Build 'as simple as possible' perceptron.
Data: KLearn digits library which is a library of 8x8 hand written images.
Process: One-layer perceptron with back propagation chosen as the simplest possible implementation of a 'Neural Network'. The maths for calculating synaptic weights, sigmoid, and errors are all written from scratch and contained in the file in order ot help give a complete understanding of how a perceptron works.
Credit: Inspired and helped by Polycode (https://www.youtube.com/watch?v=kft1AJ9WVDk)
In order to verify the success rate of the perceptron, multiple variables were chosen and the accuracy compared.
Example: training digits: 16 '3' digits and 16 'non-3' digits training cycles: tested at 2,000 , 20,000 , and 200,000 cycles testing digits: 5 '3' digits and 5 'non-3' digits. Classification accuracy measured as percent accurate classifications.
Accuracy Results: 2 cycles: 50% 20 cycles: 50% 200 cycles: 60% 2000 cycles: 50% 20000 cycles: 60% 200K cycles: 60%