Build a No-code AI Model: Teachable Machine and Machine Learning for Kids

Hello students, welcome to Kwickprep. Can you build your own AI model without writing a single line of code? Yes, you can, in a web browser. Today we collect training examples, train a model, test and improve it, and find out what the model actually learned.

Before we start, let us learn four words. A model is a program that learns from examples instead of fixed rules. Training means showing the model many examples so it can learn. A class, also called a label, is the name of a group, like Apple or Banana. A no-code tool lets you build a model by clicking and dragging, without typing any code.

We will look at two free tools. Teachable Machine is made by Google, and it learns from images, sounds and body poses. Machine Learning for Kids was made by Dale Lane, and it learns from four kinds of data: text, image, number and sound. Both run in a web browser and are free, so a school computer lab is enough.

Every no-code project follows the same steps. First, we collect examples for each class. Then we train the model with those examples. Next, we test it with new data it has never seen. Now the question, is the model good enough? If yes, we use it or export it. If no, we improve the data by adding better examples, and then train again.

Let us build a mask detector in Teachable Machine, using an Image Project. First, create one class for each group, like Mask and No Mask. Add many examples to each class, about fifty or more. You can hold a button to capture webcam photos, or upload saved pictures. Keep the classes balanced, which means roughly the same number of examples in each.

The examples must also be varied, not all the same. Use different people, of different ages, and ask each person before you photograph them. Take photos in bright light and in dim light. Change the angle and the distance from the camera. Change the background too, so the model does not learn the wall behind you. And check that every example sits in the correct class, because a wrong label teaches a wrong lesson.

Now we train. In Teachable Machine, click the button Train Model and wait a few seconds. During training, the model looks for patterns that separate one class from another. Keep the browser tab open, because the training runs inside your own browser. Under Advanced, you can see settings like epochs, which means how many times the model goes through all the examples.

Machine Learning for Kids works in a similar way. First, make a project and choose a type, like text. Second, add labels for your classes, like kind and rude. Third, add examples by typing sentences, such as, you are smart. Fourth, go to the Learn and Test page and click the train button. Fifth, use the model inside Scratch to build a game or a chatbot.

Testing means checking the model with new examples it has never seen. Riya wears a mask, and the model says Mask with ninety six percent confidence, which is correct. Aman has no mask, and the model says No Mask, also correct. When a mask is pulled below the nose, the model still says Mask, which is wrong. In a dark room, the confidence is only fifty five percent, so the model is not sure.

Pause and predict. The model fails when a mask is pulled below the nose. What should you do to fix it? Add many photos of pulled down masks to the No Mask class, and then train again.

Here are the main ways to improve a model. Add examples of exactly the cases it got wrong. Add more variety in light, angles and people. Balance the classes, so one class does not have far more examples. Look for wrong labels and fix them. Then retrain the model and test it again, because improving is a loop.

So what did the model actually learn? It learned patterns of pixels, which are the tiny coloured dots in a picture, not the meaning. It does not understand what a mask is or why people wear one. It may learn a shortcut, like a blue wall in every Mask photo. And it only knows what its examples showed, nothing more.

Here is a real kind of mistake. Suppose every Mask photo was taken in the computer lab. And every No Mask photo was taken at home. The model may learn the background instead of the face, so anyone in the lab gets called Mask. The fix is to take photos of both classes in many places, so the model must look at the face.

This teaches an important idea about AI. The data decides what a model learns, so the examples must be fair. If a group of people is missing from the data, the model works poorly for them, and this is called bias. Never collect photos or voices of others without their permission. And remember, a model can be wrong, so a human should check important decisions.

Let us revise what we learned today. Collect many examples for each class, varied and balanced. Train the model in the browser with Teachable Machine or Machine Learning for Kids. Test it with new data and look at the confidence. Improve it by adding the cases it got wrong, then retrain. And remember, a model learns patterns, not meaning. Try building your own model this week.