This repository contains a simple feedfoward neural network (random weight initialization and no regularization) that classifies XOR. The Jupyter notebook contains the XOR_Net class and demonstrates its use with a successful choice of hyperparamters in subsequent cells. There is also a Python file containing just the XOR_Net class.
The architecture of the network is as follows:
- Input layer: 2 input neurons
- Hidden layer: 2 hidden neurons
- Output layer: 1 output neuron
The network implements cross-entropy cost with sigmoid activation. Activation of the output neuron determines the binary classification (<0.5 is 0, >=0.5 is 1). Training of the network is done by gradient descent using backpropagation, which is implemented from scratch using NumPy.
The code for this network is based on the implementation in the following online text. Michael A. Nielsen, "Neural Networks and Deep Learning", Determination Press 2015.
The open-source code in the online text can also be found in Michael Nielsen's own online repository.