1The photo of the cat
One sunny morning, Luna found a photo of a cat. She turned to her robot friend Pixi with a big question.Luna: Pixi, can a computer see this cat the way I can?
Pixi: Beep boop! Not quite. A computer doesn't see a picture. It sees a giant grid of tiny squares called pixels (that's where my name comes from!), and every square is just a number.
👆 Try it: paint with numbers
0 = white paper · 9 = darkest ink
Tap any square to change its number. The computer's whole "picture" is just this table!
Prof. Spark: Exactly! A CNN is a computer brain that learns to look at all those numbers and find the cat hiding inside them.
2The magic magnifying lens
Deep in the forest, Professor Spark held up a very special magnifying glass.Luna: But how does it find anything in all those numbers?
Prof. Spark: With a magic lens! It is a tiny 3×3 window that slides over the picture, one step at a time. At every spot it asks, "Is my special pattern here?" and writes down one number.
Pixi: Beep! Let me try it. Slide the lens below and watch what I find!
👆 Try it: slide the lens
The picture
(a dark square)
(a dark square)
The lens
(its 9 numbers)
(its 9 numbers)
What the lens found
light → dark edgedark → light edgenothing here
Prof. Spark: Grown-ups call the lens a filter. Our new little picture of "where the edges are" is called a feature map.
3The shrinking spell
As the sun set, Professor Spark showed them a tidy-up trick.Prof. Spark: Too many numbers make a computer sleepy. So we shrink! Look at each 2×2 box and keep only the loudest (biggest) number. That's called pooling.
Luna: So we keep the important clues and throw away the extra?
👆 Try it: tap a "?" to shrink a box
Before (16 numbers)
➜
After (only 4!)
Tap a question mark to see which number wins its box.
Pixi: Yes! The picture gets smaller, but the best clues survive. Faster and smarter, beep!
4The forest of layers
A real CNN stacks lots of lenses and shrinking spells, like climbing a tree one branch at a time.Prof. Spark: Early layers spot simple things. Each next layer combines the clues into bigger ideas. Climb up and see!
👆 Try it: climb the layers
Luna: So it goes from lines, to shapes, to ears and whiskers, to "That's a cat!"
Prof. Spark: Right! The last layer works like a voting booth. It's called a fully connected layer, and it says how sure it is.
5Practice makes pawsome
But Pixi wasn't built knowing cats. At first, its lens numbers were totally random!Luna: Pixi, how did you learn to spot cats?
Pixi: Practice! I guess, someone tells me the real answer, and I make tiny tweaks to my lens numbers. Then again, and again, with thousands of photos.
👆 Try it: train Pixi
Prof. Spark: After enough practice, Pixi can spot cats it has never seen before. That's what learning means for a CNN!
6The CNN family tree
On the hill stood a giant tree, and every lantern on it was a different kind of CNN.Prof. Spark: All CNNs use lenses, shrinking and layers. But the family has lots of cousins with different superpowers.
Pixi: Beep boop! Some read pictures, some listen to music, some watch videos. Tap a lantern to meet them!
👆 Try it: meet the cousins
7Professor Spark's moonlight quiz
Under the stars, Professor Spark had four questions to earn a Junior Pixel Detective badge.You earned the Junior Pixel Detective badge!
For grown-ups: story words vs. real words
| In the story | Real term |
|---|---|
| Pixels as numbers | An image is a matrix of intensity values (three matrices for colour: R, G, B) |
| Magic lens | Filter or kernel; sliding it over the image is the convolution operation |
| What the lens found | Feature map (activation map) |
| Shrinking spell | Max pooling (downsampling) |
| Forest of layers | Stacked convolution and pooling layers; early layers learn edges, later ones parts and objects |
| Voting booth | Fully connected layer plus softmax, which gives class probabilities |
| Tiny tweaks | Backpropagation with gradient descent, adjusting filter weights to reduce the error |
| Shortcut bridges | Residual (skip) connections |