Week 1 · Day 3

From Pixels to a Guess

Finding edges is cool — but how does an AI turn a picture into an actual answer like "that's a triangle"? Today you draw, and a real recognizer squints at your doodle, turns it into numbers, and makes its best guess — with confidence bars you can watch.

↓ scroll to start playing
🏷️ The pro wordGiving a whole image one label is image classification. The confidence bars are a probability distribution (made by a softmax). Guessing by comparing to remembered examples is k-Nearest-Neighbors (kNN).
✏️ Toy 1 · Draw & Guess

Draw a shape — I'll guess it

🎯 Your mission: draw a Circle ○

Draw with your mouse or finger. Big and clear works best!

🤖 How sure I am:

Draw something and hit Guess! 👆
🔢 Peek Inside · What I Actually See

Your drawing becomes a few numbers

I can't understand a giant messy picture. So first I shrink your drawing down to a tiny grid of numbers (bright = you drew there, dark = you didn't). Then I compare that grid to shapes I remember. The closest match wins!

your drawing
🤖 what I see (14×14 numbers)
?
my guess
Big idea: AI turns messy real-world stuff into a small set of numbers (called features), then compares. That's the whole game: picture → numbers → compare → answer.
😈 Break It · Make Me Unsure

Can you fool the AI?

AI isn't magic — it can be wrong, and a good AI knows when it's unsure. Go back up and draw a messy scribble, or a shape halfway between two others. If my top bar drops below 45%, you've earned the Uncertainty Detective badge.

Why this matters: when confidence is low, a smart system says "I'm not sure" instead of guessing wildly. That's how self-driving cars and doctors' AIs stay safe. 🛡️
🏅 Boss Build · Shape Champion

Teach me all 5 shapes

Your mission scoreboard. Get the AI to correctly recognize each shape at least once. Draw them up in Toy 1 — this board fills in as you go.

Coming tomorrow (Day 4): right now I only know shapes someone taught me in advance. Tomorrow, YOU become the teacher — you'll train your own AI from scratch using your webcam. 🎥
🔬 Real Model Lab

A real AI names what it sees

The guesser above was a warm-up. This is ResNet-50 — a real image classifier that learned from over a million photos — running right on your device. Point your camera at a toy, a pet, a cup... or pick a photo. It gives its top guesses with confidence, just like the pros. (~25 MB one-time download; then it's cached.)

not loaded
🚀 Beyond the Basics

How the pros really do it

The compare-to-examples trick (kNN) is a real algorithm - but big products use trained networks:

🟢

kNN / template match (what you built)

✅ No training, easy. ⚠️ Slow with lots of data; weak on messy images.

🏗️

CNN classifiers

Trained nets like ResNet, MobileNet, EfficientNet. ✅ Accurate and fast. ⚠️ Need data + training.

🧠

Foundation models (AI)

Huge pretrained models like CLIP classify almost anything - even with words. ✅ Super flexible. ⚠️ Big.

🏷️ Pro names to look upimage classification · softmax · kNN · ResNet/MobileNet · CLIP.

🎉 Day 3 Complete!

You saw how recognition really works: picture → shrink to numbers → compare to examples → pick the best guess, and how AI shows how sure it is.

↺ Back to all days
You've finished the first 3 days! You now understand pixels, edges, and recognition — the core of computer vision. Next up: Day 4 — Train Your Own AI with your webcam. 🎥
AI Adventure · Week 1 · Day 3 — made for young inventors 🚀