Here's a trap that fools even grown-up engineers: if you test your AI on the same examples it practiced with, it looks like a genius — but it just memorized the answers. Real builders split their data into three piles. Let's see why.
You have 12 photos to teach a 🐱-vs-🐶 AI. Click each photo to move it between piles: Practice (train), Tune (validation), and Final Exam (test). Then grade your AI two ways and see the trick!
Click a photo to cycle: — → Practice → Tune → Final Exam → —
Every serious AI is judged on data it has never seen. The three piles each have a job:
The examples the AI actually learns from. Usually the biggest pile.
Used to adjust settings and pick the best version — without touching the exam.
Locked away until the very end. Your one honest grade. Peek at it and you've cheated!
One split is the basic version. Pros make it more robust:
✅ Simple, honest. ⚠️ Your score depends a bit on the luck of the split.
Rotate which slice is the test set and average - a much more reliable score.
Keep class balance in each split, and never let the same photo sneak into train AND test.
You learned the first rule of every real AI project: split your data, and only trust the score from data the AI has never seen.