Draw a digit (0-9) and click "Recognize".
Visualizations and analysis available here.
I trained a neural network using very naive methods on the free MNIST dataset of labeled handwritten digits The dataset itself is split between training and testing data.
Here is the repo for training and testing the model checkpointed at the version used in this page: https://github.com/xkjjx/basic-digit-recognition/tree/simple-version-checkpoint-1.
As you can easily tell, the model is extremely confident and not very accurate. It is likely heavily overfitted to the training data. Despite that, the model scores highly in the test dataset as well, scoring an accuracy of 98.03%. I ran 100 epochs 3 separate times for a total of 300 epochs.