The 42nd Period | STEM Education Hub

🎯The Challenge

read this first

Choose a small, real need at your school and build an image classifier for it with the PLTW Teachable Machine. Then do what most demos skip: test it on images it has never seen, count exactly how it fails, and write a model card, a one-page report a stranger could use to decide whether to rely on it. The deep dive covers model cards, collecting data, and failure analysis. Open the Deep Dive →

Project rules

2 or 3 classes. Each one is a thing (an object, a plant, a tray), not a person.
No faces, no photos of identifiable people, and nothing that guesses a person's identity, feelings, or health.
At least 25 training images and 10 test images per class. Test images are kept in a separate folder and never used for training.
Images are ones you took yourself, or ones whose license allows classroom use. Ask before photographing anyone's belongings.
A wrong answer must be low stakes. No medical, safety, or security uses.
Your teacher approves the plan before you build.

0 of 15 checkpoints done · your work saves in this browser as you type

🤝0 · Team Up

10 min

Agree on how you will work before you start.

Team name

Member 1Role

Member 2Role

Member 3 (optional)Role

Data Lead: Keeps the train and test folders separate, logs where every image came from, checks consent and licenses.
Builder: Runs Teachable Machine, saves each version, writes down what changed between versions.
Tester & Recorder: Runs the test set, fills in the matrix, logs failure cases.
Everyone: helps write the model card and can explain every part of it.

How will we decide when we disagree?

How will we make sure the work is shared fairly?

What do we do if someone is stuck or absent?

Everyone has a role (below). In a team of two, split the Builder and Tester jobs between you.You wrote down your three team norms.

🧭1 · Plan

Day 1

Pick a need, name your classes, and plan how you will get and test the data. Get approval before you build.

Pick a starter need, or choose "My own idea" and make sure it passes the rules

Class 1

Class 2

Class 3 (optional)

Where will the training images come from?

How will the test images be different from the training images?

My teacher approved this plan

Your project passes every rule in the Project Rules box.You described where the training images come from and how the test images will be different.Your teacher approved the plan.

📸2 · Build

Day 1–2

Collect the images, keep training and test images in separate folders, and train.

Make two folders: train/ and test/. Each has one subfolder per class. Collect the test images at a different time or place, not as extra shots from the same session.
Upload only train/ to Teachable Machine. Keep a log of where every image came from.
Train, test a few images by eye, adjust, retrain. Save each version and note what changed. Don't use your test folder for any of this.
Phone deployment (QR code) is optional.

Training images in the final version

Class 1

Class 2

You collected at least 25 training images per class.You collected at least 10 test images per class, in a separate folder, at a different time or place from the training images, and never uploaded them to Teachable Machine for training.You trained the model and saved each version (v1, v2, …). You wrote down what you changed between versions.

🔬3 · Test

Day 2

Run every test image once, record what the model said, and look hard at where it fails.

Name at least two classes in step 1 and the matrix will appear here.

Failure cases (at least 3). Wrong answers, low-confidence answers, or failures on your hard test images.

Failure 1: which image?

What the model said (and how sure)

Why we think it happened

Failure 2: which image?

What the model said (and how sure)

Why we think it happened

Failure 3: which image?

What the model said (and how sure)

Why we think it happened

You ran every test image exactly once and filled in the confusion matrix.You tried at least 5 deliberately hard images (a new background, bad light, a partly hidden object).You logged at least 3 failure cases: images the model got wrong or was unsure about, and your best explanation of why.

🪪4 · Model Card

Day 3

Write the one-page report that tells someone else what your model is for and how far to trust it. The numbers fill themselves in from your matrix.

Model name

Version you are reporting

Intended use: who would use it, to do what?

Not for: what should nobody use it for?

Limits: where does it fail, and who or what could be left out?

Checks (computed from your numbers)

2 or 3 distinct class names. Name at least two classes.

At least 25 training images per class. Enter your class names first.

At least 10 test images per class. Enter your class names first.

✗ At least 3 failure cases. You have 0 with an image and a reason.

✗ Still empty: Model name, Intended use, Not for, Training sources, How test data differs, Limits.

Your model card · not finished yet

MODEL CARD: (untitled)  [v1]
Team: (team name) — (members)

NEED: (not chosen yet)
INTENDED USE: (missing)
NOT FOR: (missing)

CLASSES: 
TRAINING DATA: . (sources missing)
TEST DATA (never used for training):  (0 in all). (how it differs from training is missing)

RESULTS ON TEST DATA
Accuracy: n/a (0 of 0)
Confusion matrix (rows = really, columns = model said):
                  

FAILURE CASES

LIMITS AND WHO COULD BE LEFT OUT: (missing)

Every section of the model card is filled in and the checks below are green or explained.You copied the finished card and handed it in the way your teacher asked.

🗣️5 · Share

Day 3

Present to another team and review theirs.

When you review another team, ask:

Could any test image have leaked into training? How do you know?
Which mistake is worse for this job, and which one did the model make more?
Who or what is missing from the training images?

You presented your card (2 minutes: the need, the matrix, one failure).You gave another team one thing that is strong and one question about their test.

📐What Strong Work Looks Like

Part	Meets the project	Strong
Planning	A real need, 2–3 object classes, rules followed, teacher approval.	The intended use and the "not for" list show you thought about who might use it and misuse it.
Data	Counts met. Train and test kept apart. Sources logged.	Test images were collected differently on purpose, and you can say how.
Testing	Every test image run once. Matrix filled in correctly.	Hard images added. You report recall for every class, not just accuracy.
Failures	Three logged, each with a reason.	Each reason is a guess you could test, and you say what you would try.
Model card	All sections complete and consistent with the matrix.	A reader who never saw your project could decide whether to trust the model.
Teamwork	Roles and norms written; everyone can explain the card.	You can describe a disagreement and how you settled it.

🛠 Project 1.1.4 · Unit 1 of 3 · 11 Weeks