Neural Network Playground

Watch backpropagation learn

This is a real multilayer perceptron trained with stochastic gradient descent and backpropagation — no libraries, just JavaScript. The shaded background shows the network's current decision boundary; dots are the training data. Try the spiral with too few neurons to see underfitting, crank the learning rate to watch training destabilize, and reset to get new random weights. A hands-on way to build intuition for how neural networks actually learn.