Neural network playground.
Draw a digit, inspect a prediction, and teach the network. See activations, weights, and gradients change through backpropagation.
Your experiment starts here
Ready to explore?
Runs in your browser.
This small educational model is initialized using synthetic digits. Its predictions are demonstrations, not a handwriting-recognition benchmark.
Quick start
Get hands-on.
Draw a digit, select its label, and add it to the training set. Train or teach the drawing, then inspect the prediction and visualization tabs.
What to look for
- 784 input values, a configurable ReLU hidden layer, and 10 softmax outputs.
- Cross-entropy loss, backpropagation, and mini-batch SGD.
- He/Xavier initialization and a warm start using synthetic digits.