INTERACTIVE DEEP LEARNING LAB
Watch a network learn.
Train a digit classifier, sketch your own number, and explore the two coordinates the model learns for every image.
Train your model
2,400 MNIST training images · 600 held-out test images
Gold: training loss (left axis). Green: held-out accuracy (right axis).
Draw a digit
Sketch one digit or load a held-out example.
Train a model to see predictions.
Hidden-layer activity
Positive units are green. Negative units are violet.
Closest training digits
Distance in the model’s learned two-dimensional space.
Explore the embedding
Each dot is one held-out digit. Select a point to inspect it.
Axes are the two learned bottleneck units. The scale is fixed across epochs within each run.
Compare training runs
Change the hidden layer size or learning rate, then train again using the same data and seed.
| Hidden units | Learning rate | Epochs | Final train loss | Held-out accuracy |
|---|
What you’re seeing
The network learns from examples using backpropagation and stochastic gradient descent. The two-unit layer compresses each image to a point, and the output layer uses that point to predict the digit. A close neighbor can look very different if the model learned to group it there.
Softmax scores show the model’s relative preferences. Hand-drawn input can differ from MNIST, and these scores are not calibrated.