WHAT WE ARE MAKING

WHY WE ARE MAKING IT

You will choose a topic that interests you, gather real examples from the world around you, and use them to train an AI model that can recognize and classify what it sees.

To understand how machine learning by building it yourself, from collecting data to watching your model make predictions.

You will go on a scavenger hunt but instead of collecting objects, you'll collect data. Pick any topic that interests you, search real-world images across at least three categories, and use them to train an AI model that can recognize and classify what it sees. Along the way, you'll discover something every machine learning engineer knows: the data you collect is just as important as the model you build.

Activity Prompt

The Data Scavenger Hunt is part of that supervised learning workflow because the AI is being given labeled examples.

Supervised Learning

؜2nd

3rd

4th

Choose Your Topic

Scavenger Hunt

Train Your Model

Test & Reflect

ACTIVITY STEPS

1st

Pens/Markers

Pencils

Paper

Laptop

Phone

ACTIVITY MATERIALS

A classification problem is when you want an AI to sort things into distinct groups.

Instead of giving the AI rules, you show it lots of examples from each group and it figures out the patterns on its own.

Most AI you interact with every day solves classification problems like photo apps that recognize faces

CLASSIFICATION PROBLEM

1st

Choose a topic with at least 3 clearly distinct categories that your model will learn to tell apart.

Questions to guide your choice:

1. What topic do you find genuinely fascinating?
2. Can you clearly describe what makes each category look or feel different from the others?
3. Can you find at least 20–30 real-world images for each category online?

Choose Your Topic

EXAMPLE: ANIMALS

Birds

Reptiles

Mammals

Why this works: each class is visually distinct, thousands of images exist online, and the differences are clear enough for a model to learn.

Eagles, sparrows, parrots, penguins, owls

Snakes, lizards, crocodiles,

turtles, chameleons

Lions, dogs, dolphins,

elephants, whales

A machine learning model learns from examples not rules. Training data is the collection of real-world images (or sounds, or measurements) you gather to show your model what each category looks like.

The more varied and accurate your examples, the better your model will learn.

TRAINING DATA

TRAINING DATA CHECKLIST

Variety

Quantity

Balance

Accuracy

Include different angles, lighting, sizes, and backgrounds

Aim for at least 10-30 examples per category. More is usually better.

Keep the number of examples roughly equal across all categories.

Every example should clearly belong to its category. Remove anything blurry or ambiguous.

2nd

Go online or take a camera and collect 10-30 images for each category. Save them in separate, clearly labeled folders.

Where to look:

Google Images, Wikipedia, iNaturalist, Unsplash, Flickr

Tips for better data:

1. Include variety: different angles, lighting, and backgrounds
2. Avoid images where two categories appear in the same photo
3. Name folders exactly as you'll name the classes in Teachable Machine

Scavenger Hunt

EXAMPLE: ANIMAL SCAVENGER HUNT

Birds

25 Images

25 Images

25 Images

Reptiles

Mammals

Target Example: 25 images x 3 categories = 75 total

Eagles, sparrows, parrots, penguins, owls

Snakes, lizards, crocodiles,

turtles, chameleons

Lions, dogs, dolphins,

elephants, whales

Teachable Machine is a free tool by Google that lets you train a machine learning model directly in your browser without coding. You upload your examples, click Train, and it builds a neural network that can classify new images in real time.

You can test your model live using your webcam or by uploading new images, and see exactly how confident the model is about each prediction.

TEACHABLE MACHINE

3rd

Go to teachablemachine.withgoogle.com

1. Select "Image Project" then "Standard image model"
2. Create one class for each of your categories and give each class a name
3. Upload your images into the correct class
4. Click "Train Model" and wait for training to complete
5. Test your model using your webcam or by uploading new images you haven't used before

Train Your Model

to add more categorie

؜Name of category

Accuracy is how often your model gets it right. But accuracy alone doesn't tell the whole story.

You also want to look at which categories confuse your model and ask why. The mistakes your model makes often reveal what it actually learned and what it missed. A model can be right 90% of the time while consistently failing on one specific category.

MODEL ACCURACY

EXAMPLE: TEST & REFLECT

Eagle photo → "Bird"

97% confidence

Dolphin photo → "Mammal"

 89% confidence

Bat photo → "Bird"

72% confidence

Key insight: The model learned visual patterns, not definitions.

This is how all ML works: it finds shortcuts in the data you give it.

✗

✓

✓

؜Exporting the model

4th

Test & Reflect

A. Test

B. Observe

C. Document

D. Share

Find 5 new images for each category: ones that were NOT in your training data. Run each through your model.

Where does it succeed? Where does it fail?

Do you notice any patterns in the mistakes?

Screenshot your model's predictions and note your observations next to each result.

Present your topic, your model, and what you discovered to the class.