TinyML

Chapter 12. Magic Wand: Training a Model

In Chapter 11, we used a 20 KB pretrained model to interpret raw accelerometer data, using it to identify which of a set of gestures was performed. In this chapter, we show you how this model was trained, and then we talk about how it actually works.

Our wake-word and person detection models both required large amounts of data to train. This is mostly due to the complexity of the problems they were trying to solve. There are a huge number of different ways in which a person can say “yes” or “no”—think of all the variations of accent, intonation, and pitch that make someone’s voice unique. Similarly, a person can appear in an image in an infinite variety of ways; you might see their face, their whole body, or a single hand, and they could be standing in any possible pose.

So that it can accurately classify such a diversity of valid inputs, a model needs to be trained on an equally diverse set of training data. This is why our datasets for wake-word and person detection training were so large, and why training takes so long.

Our magic wand gesture recognition problem is a lot simpler. In this case, rather than trying to classify a huge range of natural voices or human appearances and poses, we’re attempting to understand the differences between three specific and deliberately selected gestures. Although there’ll be some variation in the way different people perform each gesture, we’re hoping that our users will strive to perform the gestures ...