Build a Teachable Machine Image Classifier With Your Kid

October 23, 2024 | Issue 48 Companion

Most children consume artificial intelligence through algorithms that recommend videos or filter photos, yet few understand how these systems make decisions. For busy executives, introducing children to machine learning can feel daunting, especially when faced with complex coding environments or expensive software subscriptions. However, you can demystify this technology in a single afternoon using a free, browser-based tool developed by Google. By building a custom image classifier, children aged 8 to 16 can experience the shift from traditional programming, where a human writes strict rules, to machine learning, where a computer learns from examples. This hands-on project requires no coding experience, no paid accounts, and no specialized hardware, making it an accessible way to teach the core concepts of computer vision.

The Shift From Coding to Training

To help children understand artificial intelligence, you must first explain how it differs from traditional software. In traditional programming, a software engineer writes explicit instructions: if a pixel is red, do this. In machine learning, the computer is not given rules. Instead, it is given examples and must discover the rules itself. Building an image classifier is the perfect way to demonstrate this concept. By showing the computer dozens of examples of two different objects, the child trains a neural network to recognize the subtle differences between them. This process mirrors how modern AI systems, from self-driving cars to medical imaging software, are developed in the professional world.

Step 1: Accessing the Platform

To begin, open a web browser on any laptop or desktop computer equipped with a webcam. Navigate directly to Google Teachable Machine . This platform is completely free, requires no login, and runs entirely within the browser. Because the image processing occurs locally on your computer, no webcam footage is sent to external servers, ensuring complete privacy for your family. Once on the homepage, click Get Started and select Image Project, followed by Standard Image Model. This opens the training workspace, which is divided into three clear sections: Input, Training, and Preview.

Step 2: Designing the Experiment

Before capturing data, the child must decide what the computer will learn to distinguish. For the best results, choose two household items that are distinct but share some similarities. Excellent options include a coffee mug versus a water bottle, or a thumbs-up gesture versus a thumbs-down gesture. In the workspace, you will see two default boxes labeled Class 1 and Class 2. Have the child rename these boxes to match the chosen objects, such as Mug and Bottle. This step teaches the concept of labeling, which is the foundation of supervised machine learning.

Step 3: Gathering Training Data

Now, the child will collect the training data using the webcam. Click the Webcam button under the first class. Hold the first object, such as the coffee mug, in front of the camera. Click and hold the Hold to Record button to capture images. Instruct the child to slowly rotate the object, move it closer and further, and tilt it at different angles while recording. Aim to capture between 100 and 150 images. Repeat this process for the second class using the water bottle. Explain to the child that they are creating a dataset. The variety in angles is crucial because it helps the computer understand what a mug looks like in different environments, rather than just memorizing a single static image.

Step 4: Training the Neural Network

With the dataset complete, click the Train Model button in the middle column. This process takes about ten to fifteen seconds. Warn the child not to close the browser tab while training is underway. During this phase, the browser uses the computer's graphics processor to analyze the pixels in the captured images. It looks for patterns, edges, and colors that consistently appear in the mug photos but are absent in the bottle photos. The computer adjusts its internal mathematical formulas, known as weights, until it can accurately separate the two classes based on these visual patterns.

Step 5: Testing and Correcting Bias

Once training is complete, the Preview column on the right will activate, showing a live feed from the webcam. Have the child hold up the mug and observe the real-time percentage bars at the bottom. The model should confidently identify the object as a Mug. Now, try to trick the model. Have the child hold up a different mug from the kitchen, or hold the original mug upside down. If the model struggles or misidentifies the object, explain that this is called model bias. The computer only knows what it has been shown. If all the training images of the mug featured a hand holding the handle, the computer might have mistakenly learned that a hand is part of a mug. To fix this, simply record more diverse images in the training column and click Train Model again. This iterative process of testing and refining is exactly how professional AI engineers improve their models.

Parent and Educator Sidebar

Core AI Concept: Supervised Learning This project demonstrates supervised learning, where an AI model is trained on labeled data. The neural network analyzes thousands of pixels to find mathematical correlations between the input images and the labels you provided. When you test the model, it performs inference, applying its learned patterns to new, unseen images.

Conversation Starters:

• How do you think a self-driving car uses this technology to tell the difference between a stop sign and a speed limit sign?
• If we only trained our model with red mugs, why would it fail to recognize a blue mug? How does this relate to real-world AI bias?
• What other objects in our house could we train the computer to recognize?

Action Steps Summary

• Open the Tool: Navigate to Google Teachable Machine on a computer with a webcam and select a standard image project.
• Label Classes: Rename the default classes to match two distinct household items.
• Record Samples: Capture 100 to 150 webcam images of each object from various angles and distances.
• Train and Test: Click Train Model, then test the live model in the preview window and iterate to correct any bias.