The neural network consists of an input layer (2 or 4 nodes), two hidden layers with size 12 and 6, and an output layer of size 3. Both input types are included to show the importance of feature selection in model learning. While the network should be able to learn the correct decision boundary for the square and circle datasets given just the (x, y) pairs, it sometimes struggles and takes many thousands of epochs. These datasets become relatively simple given the squared input feature dimensions.

The neural network is not optimized in any way. It is about as simple as possible while still being able to provide training functionality. The goal was not to make the fastest, most accurate classifier, but rather to make a simple visualization showing the progression of a neural network trying to learn a decision boundary. Sometimes, for the simple (x, y) input, it's interesting to see the network slowly find the correct model.

Training stops after 10,000 epochs or after the network reaches 95 percent accuracy on the training set. After every 10 epochs, the network waits 0.1 seconds before continuing to give time for the decision boundary to render. Training also stops if the canvas changes size or if a different configuration is chosen.