A live backpropagation visualizer where you train neural networks on 2D classification tasks and watch decision boundaries form in real time.

Features

• Binary Rain Background — Falling 0/1 stream in neural-net blue/amber palette
• 4 Datasets — Circle, XOR, Spiral, Linear (each requiring different network depths)
• 3 Architectures — Simple (2→4→1), Medium (2→6→4→1), Deep (2→8→6→4→1)
• Decision Boundary — 60×60 heatmap showing learned classification probability
• Network Graph — Live visualization of weights and activations during forward pass
• Loss Curve — MSE training loss plotted over epochs
• Backprop Engine — Full gradient descent with configurable learning rate
• Achievement Tasks — Earn checkmarks by reaching accuracy targets on each dataset

How to Play

1. Pick a dataset (start with Circle, finish with Spiral)
2. Pick a network architecture
3. Press ▶ TRAIN and watch the decision boundary form
4. Adjust learning rate slider to avoid oscillation
5. Challenge: achieve ≥85% on the Spiral dataset!

What You Learn

• Forward pass: weighted sums + sigmoid activation
• Backpropagation: chain rule gradient flow
• Why XOR requires hidden layers (non-linear boundaries)