You’re about to train an AI that recognizes handwritten digits. But first, a mind-bender: a computer can’t see pictures at all. It only understands numbers. So how does image AI work? Because every picture is secretly a grid of numbers — and today you’ll see that with your own eyes.

💡 In Colab. scikit-learn includes a dataset of tiny handwritten-digit images.

A dataset of handwriting

scikit-learn comes with images of handwritten digits (0–9). Load them:

from sklearn.datasets import load_digits

digits = load_digits()
print("Number of images:", len(digits.images))
print("Each image is this big:", digits.images[0].shape)

There are about 1,800 little images, each just 8 by 8 dots.

Peek inside one image

Here’s the secret. Print the raw data of the first image:

print(digits.images[0])

It’s a grid of numbers! Each number is how bright one dot (pixel) is — 0 is black, higher numbers are brighter. Squint at the grid and you can almost see the digit in the pattern of big and small numbers. That is what a computer “sees”: brightness values.

Now see it as a picture

matplotlib can turn that number grid back into an image:

import matplotlib.pyplot as plt

plt.imshow(digits.images[0], cmap="gray")
plt.title("This image is labeled: " + str(digits.target[0]))
plt.show()

A blurry little digit appears, and the title shows its label (the correct answer). digits.target holds the labels — the same features-and-label idea: the features are the 64 pixel numbers, the label is which digit it is.

Try it 🎯

1. Change digits.images[0] to digits.images[7] (and target[7]) to see a different digit.
2. Show three different digits in three cells. Can you read them?

Why this matters

To an AI, an image isn’t a picture of a “7” — it’s 64 numbers. And we already know how to teach a model from numbers (that’s exactly what we did with penguins!). So recognizing handwriting is just… another features → label problem, where the features happen to be pixels. Next lesson, you’ll train it.

Think about it 🔮

A penguin had 3 number features (bill, flipper, mass). How many number features does one 8×8 digit image have? (8 × 8 = 64 — one brightness number per pixel. It’s just a bigger pile of features.)

Fix the bug 🐞

This tries to show the image but it comes out as a wall of numbers instead of a picture. It’s printing the grid instead of drawing it:

import matplotlib.pyplot as plt
print(digits.images[3])

(To draw it, use plt.imshow(digits.images[3], cmap="gray") and plt.show() — print just shows the raw numbers.)

Your mission 🚀

Explore the digit images: display five different ones with plt.imshow, each titled with its label. For one of them, also print the raw 8×8 number grid and find the brightest pixel (the biggest number) by eye.

What you learned today

• A computer can’t see pictures — it reads numbers.
• Every image is a grid of pixel brightness values; digits.images[0] is an 8×8 grid.
• plt.imshow(..., cmap="gray") turns numbers back into a picture.
• Image recognition is just a features → label problem where the features are pixels — exactly the kind of thing you can already train a model on.

Next time: you train an AI to read handwriting. 🔢