Train Your Own Model

Learning Objectives

This lesson is where World 6 turns hands-on: students don't just learn about AI, they build a small working one, using the in-app AI Playground — a webcam-based tool where a student shows the camera examples of two (or more) things, labels them, and watches the app learn to tell them apart in real time. Across the four age bands, by the end of the lesson a student should be able to:

• Explain that AI learns from labeled examples, not from being told rules. No one writes "a cat has pointy ears" into the code — the system is shown many labeled photos and works out the pattern itself.
• Describe the basic pipeline: collect data → label it → train → test → fix mistakes → use it. Every age band walks through all five steps, in the app and in the classroom activity, in that order.
• Explain why a lopsided dataset causes lopsided predictions — if you show the model 100 examples of one category and 5 of another, it will lean toward guessing the first one, because that's mostly what it's seen.
• (Ages 8+) Test a model with examples it has never seen before, and explain why testing on the same photos used for training doesn't prove anything.
• (Ages 11+) Name at least one thing that can go wrong after a model is "done" — overfitting, an imbalanced precision/recall trade-off, or (Architect) data drift after deployment — and describe, in general terms, what fixing it looks like.

Teacher Background

You do not need any coding or machine-learning background to teach this lesson — the AI Playground does the technical work, and your job is to narrate what's happening and keep the activity focused. That said, understanding what's actually going on under the hood will make you far more convincing when a sharp student asks "but how does it actually know?"

At its core, the Playground teaches a computer to sort things into categories by example — the same basic idea (called supervised learning) behind spam filters, voice assistants recognizing wake words, and photo apps that find faces. You show it several labeled examples of Category A and several of Category B, and it works out, on its own, what distinguishes the two — without anyone writing an explicit rule like "if it has whiskers, it's a cat."

Here's the part that makes the Playground special, and it's worth understanding well: it isn't training a neural network completely from scratch, which would normally require thousands of examples and real computing power. Instead, it uses a technique called transfer learning. A large image-recognition model (called MobileNet) has already been trained, by its original creators, on millions of general photos, so it already "knows" how to notice useful visual patterns — edges, shapes, textures, colors. The Playground reuses that existing knowledge and only teaches the last, small step: given the patterns MobileNet already notices in a new photo, which of the few categories the student defined does this new photo look most similar to? Concretely, it turns each of the student's example photos into a list of numbers (the patterns MobileNet noticed), and when a new photo comes in, it compares that photo's own numbers to the stored examples and picks whichever category's examples it resembles most closely. That's why the Playground can "learn" a brand-new category from a couple of dozen webcam photos in a few seconds, instead of the hours and huge datasets a from-scratch model would need.

One nuance worth knowing before you teach this lesson: the app's in-lesson scenes describe a generic AI training process — "training rounds," an accuracy percentage climbing from 50% up to 92% as training continues. That description matches how many real training tools (including Teachable Machine's own advanced training mode) work, and it's an accurate picture of how AI training generally goes. But because the Playground itself uses the fast comparison method described above rather than a lengthy training loop, students won't see a progress bar or a slowly climbing accuracy number when they actually use it — the app compares new photos to stored examples instantly. If a student notices this and asks about it, that's a great moment to explain the difference between the general concept (which the lesson teaches) and the specific shortcut the Playground takes to make it work instantly, on a phone, for free.

The rest of the pipeline is straightforward and matches real practice closely. Labeling means telling the system what each example actually is. Balancing the dataset — roughly equal numbers of examples per category — matters because a model shown mostly one category will simply guess that category more often, the same way a person shown mostly photos of golden retrievers might start calling every dog a golden retriever. Testing on unseen data is the only fair way to check whether a model actually learned the pattern, rather than just memorized the specific photos it was shown. And when it gets something wrong, the fix is never to "delete it and start over" — it's to look at what went wrong (bad lighting? a confusing angle? too few examples of that category?), add better examples, and retrain.

Materials & Prep

• A device with a browser and a working camera — phone, tablet, Chromebook, or laptop with a webcam — plus access to aiq.ph. This is the one lesson in World 6 so far where camera access is required, not optional, so confirm ahead of time that school devices allow camera permissions in the browser (a one-time "allow camera" prompt appears the first time a student opens the Playground).
• One device per student is ideal but not required. Pairs or small groups of three work well too — one student holds objects up to the camera while a partner taps to capture and label examples, then they swap roles.
• Two (or more) simple, distinct physical things to sort. The app defaults to "Thumbs Up" vs. "Thumbs Down," which needs no supplies at all. If you want the class to sort real objects instead, pick two visually distinct classroom items per pair (a pencil vs. an eraser, a red block vs. a blue block) — nothing needs to be purchased or prepared in advance.
• No accounts, logins, or data collection are needed. The camera feed is processed on the device itself and is never uploaded anywhere — worth mentioning to students and, if asked, to parents, since privacy questions naturally come up in a "the camera is watching me" activity.
• Good lighting and a bit of open floor or desk space help — dim or backlit rooms make for messier example photos, which is itself a useful lesson (see Common Misconceptions below).

Common Misconceptions

Pick your grade's script