Machine Learning for Makers: Interactive Sensor Data Classification Based on Augmented Code Examples

Although many software libraries and hardware modules support reading data from sensors, makers of interactive systems often struggle to extract higher-level information from raw sensor data. Available general-purpose machine learning (ML) libraries remain difficult to use for non-experts. Prior research has sought to bridge this gap through domain-specific user interfaces for particular types of sensors or algorithms. Our ESP (Example-based Sensor Prediction) system introduces a more general approach in which interactive visualizations and control interfaces are dynamically generated from augmented code examples written by experts. ESP's augmented examples allow experts to write logic that guides makers through important steps such as sensor calibration, parameter tuning, and assessing signal quality and classification performance. Writing augmented examples requires additional effort. ESP leverages a fundamental dynamic of online communities: experts are often willing to invest such effort to teach and train novices. Thus support for particular sensing domains does not have to be hard-wired a priori by system authors, but can be provided later by its community of users. We illustrate ESP's flexibility by detailing pipelines for four distinct sensors and classification algorithms. We validated the usability and flexibility of our example-based approach through a one-day workshop with 11 participants.