Day 4 · Train Your Own AI

🏷️ The pro wordTeaching from labeled examples is supervised learning. Starting from a model that already knows a lot and nudging it to your task is transfer learning - the secret behind Teachable Machine. The number-fingerprint of an image is an embedding.

🎥 Step 1 · Teach Me

Show me some examples

Pick 3 moves that look really different (like: lean left, lean right, hands up). Hold each one and tap Capture a bunch of times so I get lots of examples. More examples = smarter me!

Demo move:

Your camera stays on your device — nothing is uploaded. 🛡️

🧠 What I see: I shrink each snapshot into a tiny grid of numbers (just like Day 3) and remember it.

🔮 Step 2 · Watch Me Learn

Now I'll guess your move — live

Once you've taught me all 3 moves, strike a pose and watch the bars. I compare what I see now to all your examples and pick the closest match. This is real training — you made this AI!

Teach me at least 2 moves first… ☝️

🐞 Try to break it: teach two moves that look almost the same, or give one move only 1 example. Watch me get confused! (That's tomorrow's lesson: good data vs bad data.)

🎯 Boss Build · Copy Me!

Play "Copy Me!"

The real test: I'll call out a move, and you strike it. Hold it until the bar fills. The better you trained me, the easier this is. Get 10 to become an AI Master!

Score: 0

🚀 Beyond the Basics

How the pros really do it

Your webcam trainer uses nearest-neighbor on features. Pros use the same idea, supercharged:

🟢

kNN on features (what you built)

✅ Learns from a handful of examples instantly. ⚠️ Limited accuracy.

🔁

Transfer learning

Fine-tune a pretrained CNN on your classes (Teachable Machine, MobileNet). ✅ Great results from few examples. ⚠️ Some setup.

🧠

Few-shot with big models (AI)

Modern models classify from a few examples using embeddings (CLIP). ✅ Powerful and flexible.

🏷️ Pro names to look upsupervised learning · transfer learning / fine-tuning · embeddings.

🎉 Day 4 Complete!

You didn't just use an AI — you trained one from scratch, from your own examples. That's exactly how real AIs learn: from data.

Day 5: Good Data vs Bad Data →

Tomorrow: your AI is only as smart as the examples you gave it. We'll see how bad data makes even a good AI act silly — and how to fix it. 🧪