Training a Machine Learning Model with CreateAI

CreateAI is a free, web-based tool that makes it easy to explore artificial intelligence with micro:bit. In this tutorial, you’ll learn how to use CreateAI to train a machine learning model using accelerometer data from your micro:bit. We’ll train the model to recognize different physical actions, like a wave, a thumbs-up, or jazz hands, which you can then use to control other projects.

CreateAI Dashboard

Materials

micro:bit (V2)
Power source for the micro:bit, like the CHARGE for micro:bit
micro USB cable
Computer with internet access

Instructions

A video walk-through of this tutorial is available here:

Step 1: Connect micro:bit to CreateAI

First, we need to establish a connection between our micro:bit and CreateAI.

Start something new. Connect your micro:bit and collect movement data to build a machine learning model.

Open the CreateAI dashboard and select ‘New Session’.

Screenshot of connect prompt in CreateAI's platform.

You’ll be prompted to connect your micro:bit with a micro USB cable. Follow the on-screen instructions to pair your device.

CHARGE for micro:bit strapped onto someone's wrist.

Once your program download is complete, disconnect your micro:bit from your computer and attach it to a separate power source like the CHARGE!

You will then be prompted to connect the micro:bit via Bluetooth. Match the pattern on the micro:bit’s LED display to the grid on the CreateAI window and accept the prompts to pair.

Step 2: Collect Data Samples

Once your micro:bit is wirelessly connected, you’re ready to collect some data! We’ll start by labeling different actions we want the model to recognize, then adding some sample data for what each action looks like.

Remember: CreateAI uses the micro:bit’s accelerometer to detect different actions. As you move your hand throughout the activity, you’ll see a live data graph showing its movement along the X, Y, and Z axes, which represent different directions in space.

Live data graph showing the micro:bit's movement along the X, Y, and X axes.

Complete the following steps for each action you want to train. We recommend starting with at least three actions. The more unique each action is, the more consistently your micro:bit will be able to tell the difference between them.

First action named 'waving' with a smiley face symbol.

Choose a new action (e.g. ‘waving’). Name it, then select a symbol from the dropdown that best represents it.

Follow the on-screen instructions to record a few seconds of accelerometer data for your new action. The more data you include, the more reliably your micro:bit will be able to recognize it. You’ll need to record three samples to start.

Click ‘+ Add action’ (bottom left) to repeat the process for more actions.

Step 3: Train Model

After you have collected enough data for all your actions, click the ‘Train’ button (bottom right) in CreateAI. This will begin the process of training your machine learning model and may take a few moments.

Training is when the computer learns to identify patterns in the data you provided. It’s like teaching a child what a cat is by showing them many different pictures of cats. The more data you provide for each action, the better and more accurate the model becomes at making predictions.

Step 4: Test Model

Once the model is trained, you can test it directly within CreateAI. Perform the actions you trained the model to recognize (e.g., wave your micro:bit) and observe how the model identifies the action in real-time. You’ll see a Certainty score that indicates how confident the model is in its prediction. You can adjust the Recognition Point to set the confidence level at which the model should identify an action.

Next Steps

Now that you have a model trained to recognize different actions, you can use these as inputs to trigger responses (or outputs) in MakeCode. To learn how to combine your trained model with MakeCode programs, check out the following tutorials and projects: