TinyML

Chapter 11. Magic Wand: Building an Application

So far, our example applications have worked with data that human beings can easily comprehend. We have entire areas of our brain devoted to understanding speech and vision, so it’s not difficult for us to interpret visual or audio data and form an idea of what’s going on.

A lot of data, however, is not so easily understood. Machines and their sensors generate huge streams of information that don’t map easily onto our human senses. Even when represented visually, it can be difficult for our brains to grasp the trends and patterns within the data.

For example, the two graphs presented in Figure 11-1 and Figure 11-2 show sensor data captured by mobile phones placed in the front pockets of people doing exercise. The sensor in question is an accelerometer, which measures acceleration in three dimensions (we’ll talk more about these later). The graph in Figure 11-1 shows accelerometer data for a person who is jogging, whereas the graph in Figure 11-2 shows data for the same person walking down stairs.

As you can see, it’s tough to distinguish between the two activities, even though the data represents a simple and relatable activity. Imagine trying to distinguish between the operating states of a complex industrial machine, which might have hundreds of sensors measuring all sorts of obscure properties.

It’s often possible to write handcrafted algorithms that can make sense of this type of data. For example, an expert in human gait might ...