Build a Teachable Machine Image Classifier With Your Kid

Issue 36 • May 8, 2024

Most children interact with artificial intelligence daily through video game matchmaking, video recommendations, or voice assistants, yet few understand how these systems actually learn. By building a custom image classifier together, you can demystify machine learning and show your child how computers are trained to see the world. Using a free, web-based tool, this hands-on project requires no coding experience and can be completed on any computer with a webcam.

For busy executives, finding meaningful, educational activities to share with children can be challenging. This project offers a perfect balance of high-tech engagement and low-friction setup. It moves children from passive consumers of technology to active creators. By using Google's Teachable Machine, a free tool designed to make machine learning accessible, you and your child can train a computer vision model in under fifteen minutes.

The project is ideal for children aged eight to sixteen. Younger children will enjoy the immediate feedback of seeing the computer recognize their favorite toys, while older children can dive deeper into the concepts of data bias, model accuracy, and real-world applications.

Setting Up Your Workspace

To begin, you only need a computer (Mac, Windows, or Chromebook) with a working webcam and an internet connection. No software installation or account registration is required, making this a highly secure and private activity.

• Sit down with your child and navigate to the official tool at https://teachablemachine.withgoogle.com/ .
• Click on the Get Started button on the homepage.
• Select Image Project from the list of options, and then choose Standard Image Model. This opens the training interface, which is divided into three clear sections: Input, Training, and Preview.

Step 1: Defining Your Classes

Before gathering data, you must decide what you want your computer to recognize. A highly engaging theme is distinguishing between a healthy snack (like an apple or a banana) and a sweet treat (like a cookie or a juice box).

In the first box, labeled Class 1, click the pencil icon to rename it. Let us name it "Healthy Snack."

In the second box, labeled Class 2, rename it to "Sweet Treat."

You can also add a third class by clicking Add a class and naming it "Empty Hand" or "Background." This is a crucial step because it teaches the computer what the camera looks like when no objects are present, preventing false positives.

Step 2: Gathering Training Data

Now comes the hands-on part for your child. They will act as the data scientist, gathering the examples the computer needs to learn.

Click the Webcam button in the "Healthy Snack" class box. Have your child hold the healthy snack in front of the webcam. Click and hold the Record button. Have your child slowly rotate the snack, move it closer to the camera, and move it further away. Capture at least one hundred images. This variety ensures the model learns the shape and color of the object, rather than just one specific angle.

Repeat this exact process for the "Sweet Treat" class, capturing another one hundred images of the cookie or juice box.

Finally, record about fifty images of just the empty background or your hand without any objects for the third class.

Step 3: Training the Model

With the data gathered, it is time to train your neural network. Click the Train Model button in the middle column.

Explain to your child that the computer is now looking for patterns in the images. It is analyzing colors, edges, and shapes to find the differences between the classes. Keep the browser tab open and active during this process. It usually takes less than a minute. Once completed, the Preview column on the right side of the screen will activate, turning on your webcam.

Step 4: Testing and Debugging

This is where the real learning happens. Have your child test the model by holding up the healthy snack. The preview window will show a real-time bar graph indicating how confident the model is.

Now, challenge your child to break the model. What happens if they hold up a completely different healthy snack, like a carrot, that was not in the training data? What happens if they hold up a red ball that looks like the red apple? When the model makes a mistake, explain that this is called a classification error. It happens because the training data was too limited. To fix it, you can simply record more diverse images in each class and click the Train Model button again.

Parent and Educator Sidebar

Core AI Concept: Supervised Learning and Training Bias

Supervised learning is the process of training an AI model using labeled data. The computer does not inherently know what an apple is. It only knows that the images in the "Healthy Snack" folder share certain pixel patterns. If we only train the model with red apples, it will fail to recognize a green apple. This is training bias, a major challenge in real-world AI development, from facial recognition to medical diagnostics.

Conversation Starters:

• "How does the computer know the difference between your hand and my hand?"
• "What would happen if we only trained the model with pictures of your hand, and then tried to use it on my hand?"
• "How do you think self-driving cars use this kind of technology to see stop signs?"

Action Steps Summary

• Open a web browser and go to https://teachablemachine.withgoogle.com/ .
• Create a new Standard Image Model project.
• Define three classes: Healthy Snack, Sweet Treat, and Empty Hand.
• Record at least one hundred webcam images for each object class from various angles.
• Train the model and test its accuracy using new objects.
• Discuss how training data limitations affect the model's real-world performance.

Explore more AI workflows at PromptHacker.