class 1 / positive weight / positive activationclass 0 / negative weight / negative activation
Couldn't start the 3D scene.

Backpropagation Playground: Training a Small Neural Network to Bend Decision Boundaries

Train a real multilayer perceptron live by backpropagation. Pick XOR, a disc in a ring, two moons or two spirals, change width, activation, learning rate and noise, and watch the weights, the forward-pass activations of a moving probe, the loss curve and the decision map evolve. A Blender-built network on a plinth with a display, and numerical gradient checks built in.

Made with 🟠 Blender · Three.js