Day 2 · Find the Edges

🏷️ The pro wordSliding a small grid of numbers over an image is a convolution; the grid is a kernel (or filter). Stacking layers of learned filters makes a Convolutional Neural Network (CNN) — the engine of most vision AI.

🪟 Toy 1 · The Sliding Window

Slide, multiply, add — that's it

On the left is a shape made of numbers (9 = bright, 0 = dark). A little 3×3 window slides across it. At each spot it multiplies the 9 numbers by a secret recipe and adds them up — and drops the answer on the right. Watch where the answers light up!

the picture (numbers)

edges found ✨

Press Play to start sliding…

Whoa: in the middle of the shape all the numbers are the same, so the answer is 0 (stays dark). Only at the edges, where numbers suddenly change, does the answer get big and light up. That's how a computer finds edges! Grown-ups call this sliding trick a convolution.

🎛️ Toy 2 · The Recipe Lab

Change the recipe, change the magic

The 9 numbers in the window are called the recipe (real name: kernel). Different recipes do totally different things — blur, sharpen, outline. Type your own numbers, or tap a preset, and watch a real photo change instantly.

before

→

after

🎯 Challenge: turn the photo into just outlines. Tap Edges, or build your own: a big number in the middle and negative numbers around it.

🧠 Big Picture · How AI Really Sees

Real AI stacks thousands of these

A computer-vision AI isn't magic. It slides hundreds of recipes like yours over a photo. One finds up-edges, one finds side-edges, one finds curves… stack them and it starts spotting eyes, wheels, letters, whole objects. Tap each filter to see what it grabs.

The idea: simple edge-finders → combine into shape-finders → combine into object-finders. On Day 4 you'll train an AI that learns its own recipes from examples you give it. Grown-ups call this stack a convolutional neural network (CNN).

🧪 Boss Build · Invent a Filter

Make your own magic recipe

Play mad scientist. Change the 9 numbers above until the photo looks awesome or weird, give your filter a cool name, and save it to your lab notebook.

Save your first filter to earn the Filter Inventor badge. Your saved recipes appear here.

🚀 Beyond the Basics

How the pros really do it

Hand-designing an edge filter shows the idea. Here's how it scales up:

🟢

Hand-made kernels (what you built)

Classic filters like Sobel and Canny for edges. ✅ Fast, no training. ⚠️ You must design each one; brittle.

🧩

Classic feature detectors

SIFT / HOG — clever hand-designed features that powered vision before deep learning.

🧠

CNNs and Transformers (AI)

A CNN learns thousands of filters automatically; newer Vision Transformers (ViT) use attention. ✅ Far more powerful. ⚠️ Need data + training.

🏷️ Pro names to look upconvolution · kernel/filter · Sobel/Canny · CNN · Vision Transformer (ViT).

🎉 Day 2 Complete!

You learned the #1 trick of computer vision: slide a tiny recipe across the numbers to find edges and shapes.

Day 3: From Pixels to a Guess →

Tomorrow: now that the AI can find edges and shapes, how does it turn them into an actual answer like "that's a cat"? You'll draw, and watch a real recognizer make its guess with confidence bars. 🎯