Interactive lab

Train an MNIST classifier

Tune the network, watch it learn, then test it with your own handwriting.

Ready to load MNIST.

Training split 80%

Examples

Layer sizes Input 784 and output 10 are fixed; add up to three hidden layers.

Hidden activation

Output activation

Cost function

Learning rate

Momentum

Mini-batch size

Epochs

Epoch0 / 8

Training accuracy—

Validation accuracy—

Validation loss—

Macro AUC—

Evaluation

Training and validation accuracy after each epoch.

Live inference

See how the trained model ranks every possible digit.

Draw with a mouse, pen, or finger. The image is centered and reduced to the same 28 × 28 format as MNIST.

Prediction

Train the model first