Week 10 — Training Data: Teach a Machine Yourself¶

At a glance


Date	Friday, October 23, 2026
Unit	2 — AI Unmasked
Big idea	A machine learning model improves or fails based on the examples we give it.
Goal	Every pair trains, tests, and improves a simple image classifier in Teachable Machine.
Student accounts needed	None. Teachable Machine runs without login.
Prep time	5 minutes, plus optional image-pack prep if cameras are blocked

Why this lesson looks the way it does¶

This is the anchor experience for Unit 2. The goal is not for students to "use an AI website." The goal is for them to feel the causal chain: we choose examples → the model learns patterns → the model behaves well or badly because of those choices.

Keep the scope small: two classes, obvious objects, fast testing. A perfect model is less valuable than a model that fails in an understandable way and gets better after students add better data. That is the teachable moment.

Your industry lens belongs in the debrief. In real AI projects, teams spend enormous time on data quality, edge cases, evaluation, and deciding what "good enough" means. Students can grasp that if they have just watched their own sloppy examples produce sloppy behavior.

CSTA 2026 standards¶

Concept area	Connection
Algorithms & Design	Students design a training-and-testing process for a classifier.
Programming	Students observe that the model changes behavior without them writing explicit rules.
Data & Analysis	Students collect, label, test, and revise training data.
Systems & Security	Students use a browser-based AI tool without accounts and discuss camera/upload constraints.
Computing & Society	Students connect model reliability to choices made by people collecting data.

"I can…" objectives¶

By the end of today, students can say:

I can train a simple image classifier with labeled examples.
I can test whether my model works on new examples.
I can improve a model by adding better training data.
I can explain why training data quality matters.

Teacher prep — the 5-minute version¶

Open https://teachablemachine.withgoogle.com on the projector.
Click through once before class: Get Started → Image Project → Standard image model.
Write on the board:

Last Friday: AI = examples → patterns → predictions
Today: We choose the examples.
Also write the class challenge:

Train: Class A vs. Class B
Test: new examples only
Improve: add missing examples
Have a camera fallback ready. If Chromebook camera access is blocked, students use Upload instead of Webcam. The fastest fallback is a teacher-prepared folder of simple images: 10–15 examples each of two classroom categories, such as pencil / not pencil, book / water bottle, or thumbs up / open hand. Put the folder somewhere students can access without a new account, or project your copy and run one shared class model.

Camera access may be blocked

Teachable Machine supports image upload. Do not let the lesson depend on webcams. If uploads are also hard on student devices, make it a whole-class model: students bring objects to the projector while you add examples from the teacher machine.

Minute-by-minute run sheet¶

0:00–0:04 — Anchor¶

"Last Friday we defined AI as examples, patterns, and predictions. Today you are in charge of the examples. If the machine learns badly, we will not blame the machine first. We will inspect what we taught it."

Call back to Week 1:

"Remember Quick Draw: It learned from us. Today a model is going to learn from this room."

0:04–0:12 — Mini-lesson: the loop¶

Project Teachable Machine and narrate only the essential steps.

"This tool has three jobs. First, we create classes — the labels we want it to recognize. Second, we give examples for each class. Third, we train and test."

Draw this loop on the board:

Collect examples → Train → Test on new examples → Improve the examples

Say:

"Testing on the same examples you trained with is like giving yourself the answer key. The real test is a new pencil, a new angle, a new hand, a new lighting condition."

Model a two-class setup quickly: thumbs up and open hand, or book and water bottle. Do not overbuild the teacher model; students need time.

0:12–0:36 — Paired work: train your classifier¶

Pairs use Driver/Navigator. Driver controls the Chromebook; Navigator watches for data quality.

Student steps:

Go to https://teachablemachine.withgoogle.com.
Choose Get Started → Image Project → Standard image model.
Rename the two classes.
Add at least 20 examples per class using Webcam or Upload.
Click Train Model.
Test with new examples that were not in the training set.
Add 5–10 better examples where it fails, then retrain.

Say:

"Your first model does not need to be good. In fact, I hope it makes one interesting mistake. Your job is to catch the mistake and improve the data."

Circulate with these questions:

"Are all your examples from the same angle?"
"Did you accidentally train the background instead of the object?"
"What new example would be fair but harder?"
"What changed after retraining?"

Good pair categories

Choose categories students can produce quickly and safely at desks: closed fist/open hand, pencil/eraser, book/water bottle, index card face-up/face-down, two colored paper shapes. Avoid faces as the main category; it invites privacy and bias issues too early.

Ask for failures first.

"Who has a model that made a funny or surprising mistake? That is the one we want to see."

Project one pair if possible. Ask:

"What did the model probably learn that you did not mean to teach it?"

Then land the professional point:

"A lot of real AI work looks like this. Not magic. Not just code. Careful examples, careful tests, and honest measurement of where the system fails."

0:42–0:45 — Teaser + prompt¶

"Next Friday we are going to do something slightly unfair on purpose. We will train a model with lopsided data and watch it fail. That failure has a name: bias."

Tiered challenges¶

🟢 Starter (grades 5–6): Train two clear classes and explain one mistake your model made.
🟡 Builder (grades 7–8): Improve your model by adding examples from a new angle, distance, or lighting condition. Explain whether it helped.
🔴 Extender (grades 9–10): Design a fair test set: five new examples per class that are not copies of the training data. Report accuracy before and after retraining.

"What did your model learn that you meant to teach it — and what might it have learned by accident?"

🤔 Think About This¶

Notice one machine learning mistake this week. Autocorrect, recommendations, search, captions, anything.

Ask: what example might have been missing from its training data?

🔌 Unplugged Plan B — if the internet is down¶

Run a paper classifier. It still teaches the training loop.

Time	Activity
0:00–0:04	Same anchor: today we choose the examples.
0:04–0:10	Explain classes and training examples using two piles of cards: cats and dogs, or safe shapes and not safe shapes.
0:10–0:25	Pairs create a training set for a secret two-class rule using index cards. Example: "Class A has rounded edges; Class B has sharp edges." They make 8 examples per class.
0:25–0:35	Trade with another pair. The receiving pair infers the pattern and classifies 6 new test cards.
0:35–0:41	Original pair reveals the rule. Discuss which training examples helped and which confused the classifier.
0:41–0:45	Same teaser and Think About This.

Common problems & fixes¶

Problem	Fix
Teachable Machine asks for camera permission and it is blocked	Use Upload. If students lack image files, run one whole-class model from the teacher device.
Students train on only 3–4 examples	Set a visible minimum: 20 examples per class before Train.
Model seems perfect immediately	Make the test harder: new angle, new person, new background, different lighting.
Model learns the background	Have students move the object or hand to a different part of the frame and add examples.
Students want three or four classes	Keep most pairs at two. Extenders may add a third only after a working two-class model.
Tool runs slowly	Reduce classes to two and examples to 15 per class; pair students on fewer machines if needed.

Vocabulary¶

Class — a category the model is trying to recognize
Label — the name attached to a training example
Model — the learned pattern the computer uses to make predictions
Training — the process of learning from examples
Test set — new examples used to check whether the model works

Reflection — fill in after class¶

Did cameras work, or should Week 11 assume upload/teacher-led?
Which pairs discovered background or lighting problems?
Did students test with new examples or only training examples?
Which failure would be useful to reference next week?

Next week: Week 11 — Bias: When the Data Is Lopsided