01 · Draw something
0–9Draw one symbol. Predictions update automatically. Tap the correct answer to teach the network.
Actual input to the network28 × 28 greyscaleAutomatically fitted and centred0 active pixels
Tip: try a built-in example before drawing. Some handwritten letters look identical in different cases.
02 · Neural network
Ready for input Positive weight Negative weight
Tap a neuron to inspect
All 784 input neurons are calculated. Selected neurons are displayed.
03 · Prediction and feedback
Click to teachRecognised character
–
Draw a character to begin
Teach the correct answer
Tap the correct answer, even when the prediction is right. Each tap performs backpropagation.
Uppercase and lowercase are separate choices. Case matters.
Draw something or try an example, then tap the correct answer to train it.
The digit recogniser uses real handwritten samples. Freehand accuracy can vary; corrections help adapt it.
Explore · What is the neural network calculating?
Live neuron inspectionStep through each layer, click a visible neuron, and inspect its activation, weights, bias and the contributions from connected neurons.
Train · Backpropagation and saved examples
Runs in your browserQuick training happens whenever you click the correct answer. Each click adds the labelled drawing and changes the network weights.
0 saved drawingsLast loss: —
No examples collected yet.
Training progress: Cross-entropy loss measures how strongly the model disagrees with the correct answer. For the same example, lower is better. One example can improve without improving overall accuracy.
Original pre-trained weights loaded.
Model changes are saved in this browser when local storage is available. Export the JSON file to keep a backup or move your trained network to another computer.