Build, visualize, and train neural networks entirely in your browser — no server, no dependencies, no setup.
- Drag & drop layers and neurons on an infinite canvas
- Auto-connect with 6 connection modes or connect manually
- Undo/Redo, Auto Layout, Snap to Grid
- Semantic zoom: full detail → collapsed → minimal
- Context menus, keyboard shortcuts, minimap navigation
- Dense — fully connected layers
- Dropout — inverted dropout with configurable rate (0–0.9), visual overlay during training
- Batch Normalization — post-linear, pre-activation; training uses batch stats, inference uses running EMA
- C backend compiled to WebAssembly — runs entirely in a Web Worker (UI never freezes)
- 4 optimizers: SGD, SGD + Momentum, Adam, RMSprop
- 9 activations: ReLU, Sigmoid, Tanh, Softmax, LeakyReLU, ELU, GELU, Swish, Linear
- 5 loss functions: MSE, Binary CrossEntropy, Categorical CrossEntropy, MAE, Huber
- Early Stopping with configurable patience and min delta
- Configurable learning rate, batch size, epochs, train/val/test split
- Real-time loss and accuracy charts with validation loss tracking
- Weight-colored connections (blue = positive, red = negative)
- Neuron activation heatmap
- Decision boundary plot (2D classification)
- Forward pass animation
- Tooltips on hover (weights, biases, activations)
- Upload CSV/JSON, paste data, or load from URL
- Synthetic generators: Moons, Circles, Spiral, Gaussian Blobs, Checkerboard, XOR, Iris, Regression, Autoencoder
- Manual dataset editor
- Per-column role assignment (feature/target) and normalization
- Classifier (Iris) · Deep · Wide · Autoencoder · Binary (XOR) · Regression
- Dropout Regularization · Batch Normalization
- Custom (define layer sizes interactively)
- Network topology as JSON
- Trained weights as JSON
- Canvas snapshot as PNG
| Component | Technology |
|---|---|
| Frontend | HTML5 Canvas, Vanilla JS, CSS |
| Backend | C99 → WebAssembly (Emscripten) |
| Training | Web Worker (off main thread) |
| Charts | Chart.js 4.4.7 |
| Icons | VS Code Codicons |
Requires Emscripten:
cd backend
make clean && makeOutputs build/vnns.js and build/vnns.wasm.
Serve the root directory with any static file server:
npx serve .
# or
python -m http.server 8000Open http://localhost:8000.
MIT