Inside the experiment
What a little network sees
A handwritten digit, 784 pixels, and the useful gap between what we draw and what a model understands.
Sample post. An illustrative notebook entry, not a product announcement or research result.
Sample notebook entry for the blog layout. This is an illustrative explanation, not a new research result or product announcement.
Draw a seven and you see a seven. A small classifier sees something less familiar: a grid of numbers, each describing how bright one pixel is. That difference is a good place to start an experiment.
Make the invisible part visible
In Digit lab, the drawing is only half the story. The input preview shows the smaller image that reaches the network. A sweeping stroke can become a handful of bright squares; a subtle hook can disappear.
Putting that preview beside the prediction makes mistakes more interesting. Instead of asking only “did it get the answer right?”, we can ask what information survived the trip from our hand to the model.
A score is not a promise
Accuracy on held-out MNIST images describes performance on that test set. It does not guarantee the same result for a finger drawing, a messy canvas, or a digit with an unfamiliar shape. Those are different inputs.
Try a real held-out example, then draw your own version. Compare the model input, not just the winning label. The difference is often more useful than the score.
Keep the loop small
A canvas, a preview, and ten probabilities are enough for a surprisingly rich conversation with a tiny network. Change one thing, look closely, and try again. The point is not to make the model seem magical. It is to make its limits easier to see.