An Artificial Neural Network

A Feed Forward Artificial Neural Network using numpy with variable number of hidden layers. This is tested on the MNIST dataset.

Install and run

• git clone https://github.com/ryan-tabar/An-Artificial-Neural-Network
• pip install -r requirements.txt
• python MyANN.py

Motivation

I wanted to have a go at creating an Artificial Neural Network (ANN) from first principles to better understand how it works. Therefore, I am not using machine learning libraries such as Tensorflow or PyTorch. This is so I can learn the underlying maths thats involved.

How to use the Neural Network class

The class takes 4 __init__ arguments:

• my_ANN = NeuralNetwork(input_nodes=784, hidden_nodes=38, output_nodes=10, hidden_layers=2)

To train the network is to call upon the .train method:

• my_ANN.train(inputs, targets, learn_rate=0.1, epochs=1)
• inputs and targets must be a numpy array
• epochs= number of times to train the same training set

To use the trained network is to call upon the .feed_forward method:

• prediction = my_ANN.feed_forward(inputs)
• inputs must be a numpy array

Tests

The following are the parameters that have made the ANN achieve an accuracy of 92-93% on the 10,000 test images:

• input_nodes = 784
• hidden_nodes = 38
• output_nodes = 10
• hidden_layers = 2
• learn_rate = 0.1
• epochs = 1 (all 60,000 training images once)

The Maths involved

All the maths I’ve worked out can be found in the word document in the Maths folder or click the following: https://github.com/ryan-tabar/An-Artificial-Neural-Network/blob/master/Maths/NeuralNetworkEquations.docx

Here are the screeenshots of that document: