DigiSim

DIGIT/ LAB

Runs on your device
EN
AN INTERACTIVE MACHINE LEARNING LAB

See how learning happens.

Start with your handwriting. Follow the numbers. Teach a real neural network.

UNTRAINEDCheckpoint 0
02INSIDE THE NETWORK
784 → 32 → 16 → 10
Checkpoint 0 · frozen
Drag to orbit·Shift-drag to pan·Scroll / pinch to zoom
Preparing volumetric geometry…
3D SPACEHidden neurons occupy width, heightand depth. Spacing changes the view, never the model. The input image correctly remains a 2D plane.
Pixels: 0 → 1ReLU activationSigned output score
Select a neuron to inspect its incoming connections. All 25,760 connections are used in computation.
03FOLLOW ONE EXAMPLE
START WITH A QUESTION

Can it read your handwriting yet?

The network begins with random weights, not knowledge of digits. Draw a digit or select an example, then follow the numbers behind its prediction.

Inference reads the weights. Only an explicit training action changes them.
MAKE A MISTAKE. CHANGE A WEIGHT. TRY AGAIN.

Learning is a measurable change.

0completed epochs
0examples processed
0SGD updates
—last mini-batch mean loss
—validation accuracy
Cross-entropy lossLower is better
Accuracy on labeled examplesHigher is better
TrainingValidation · each point evaluates the entire indicated split at a fixed checkpoint.
Inspect every chart point and its sample count

Each row evaluates the full indicated subset at one fixed checkpoint. Loss is the arithmetic mean across those examples; accuracy is correct / n. Changing the training-data amount changes the training curve’s sample count, shown below. Validation always uses the same held-out set.

Try an experiment: will bigger learning steps learn faster?

Reset, train at η = 0.07, and inspect the validation curve. Reset to the identical initial checkpoint and repeat at η = 2.0. Compare measured results; improvement is not promised. To explore memorization, use fewer training examples and compare training performance against validation. Reserve the test split for your final evaluation.

DIGIT / LAB/Pixels → predictions → mistakes → learning.Real data. Real gradients. Entirely in your browser.
YOUR FIRST EXPERIMENT

Teach it. Then challenge it.

Draw a digit or click a labeled example. The network initially has random weights, so a convincing-looking prediction may be wrong.

1. Follow one prediction

UseStep through prediction. See 784 pixel values become 32 activations, then 16, then ten scores and probabilities. Select any sphere or use the layer/index controls to inspect every incoming term.

2. Follow one training step

Choose a training example, or recognize your drawing and supply its intended digit. UseTeach this exampleand advance through loss, output gradients, hidden gradients and the proposed update. OnlyApply SGD updatechanges the weights.

3. Train on more handwriting

UseStart real training. A mini-batch combines gradients from several labeled examples. Pause at a batch boundary to inspect an immutable snapshot. The Update inspector shows the real batch-mean gradient, not the gradient of the displayed sample.

4. Test something new

Evaluate on the test split; inspect the confusion matrix and mistakes. Then try your own drawing. Canvas handwriting may differ substantially from this small training dataset; even a high probability can be wrong.

Data and privacy

The bundled data is UCI Optical Digits, not MNIST: 1,797 genuine 8×8 examples normalized into 28×28 input planes. The optional MNIST importer downloads official IDX archives from the Museum’s verified CVDF copies. Drawings and computations stay in this browser. Closing/reloading resets the lab; export weights or measurements to preserve them.

Camera and accessibility

Drag to orbit; Shift-drag or right-drag to pan; scroll or pinch to zoom. Focus the scene and use arrow keys, +/− and Home. All neurons can be selected without the 3D view. Instructional playback can be paused and respects reduced-motion preferences.