Pixels are just numbers
Every image in this dataset is an 8x8 grid of numbers. Each number (0-16) represents how dark that pixel is. That's all a "digit image" is — 64 numbers.
Your task: draw a digit on the left grid and try to match the real one on the right. Click and drag to paint. Use the slider to control how dark your brush is.
Your drawing
Click & drag to draw
← match →
Real digit 7 (from dataset)
This is what the model trains on
What the model actually sees
To a model, your drawing isn't a picture — it's a flat row of 64 numbers. Every change you make above updates this array in real time:
Loading...
Loading...
Loading...
Think Deeper
Try this:
In the Explore tab, edit the grid to draw a 3. Then change just ONE pixel. Does it still look like a 3?
A single pixel change barely matters — the model looks at the whole pattern. This is why ML is robust to small noise but vulnerable to adversarial examples that change many pixels in a coordinated way.
Cybersecurity tie-in: In malware analysis, features aren't pixels — they're things like
file entropy, import counts, and section sizes. But the principle is identical: every feature is just
a number, and every sample is just a row of numbers.