Typically, the first iteration produces a model that does not work flawlessly. The most common issue is false detections, where the model predicts one of the target classes when the subject is performing a different action. To address this:
- Test the model on device.
- Remember which unintended movements result in false detections.
- Collect raw sensor data for those movements and add them to the unknown class.
- Retrain the model.
- Test again.
A few iterations of this process are usually necessary to achieve high accuracy and a low rate of false detections.