A neural network that learns to read your handwriting
The video explains it step by step in two silent minutes. Meanwhile the same network is training next to it, live, on 1,797 real handwritten digits. When the bar fills, draw one.
Draw big, filling the box.
Getting the training data…
How it learned to read your handwriting, not just the dataset's
Three things, none of them a bigger network. Every training digit is shifted, rotated and rescaled afresh on each pass, so thirty epochs are thirty different views. Your drawing is framed on its centre of mass, the way the training images were. And each prediction is averaged over seven small variations of your drawing, which stops it flickering mid-stroke.
Measured against the previous version of this demo: 93.3% → 96.7% on clean held-out digits, and 44.8% → 82.3% on the same digits shifted and rotated the way a hand would. Press Retrain to re-split the data and see the number move by a point or two: the honest answer to "how good is this model" is a range.