A neural network that uses ReLU instead of sigmoid. This makes training, even with random hill climbing, much faster. I also made a few tweaks to automatically stop training after it gets the correct output for every data set. And I further optimized the data set mode by training it based on the total loss instead of the individual losses. Press space twice to start training on the XOR function. Default neural network has 2 inputs, 1 output, and 2 hidden layers with 3 neurons each. Training does not use backpropagation, as that is too complicated for me to understand. Instead it uses a form of stochastic search called random hill climbing.
