A Teaching Language for Building Object Detection Models

ACM conference on Designing Interactive Systems (DIS) |

Object detection is a key application of machine learning. Currently, these detector models rely on deep networks that offer model builders limited agency over model construction, refinement and maintenance. Human-centered approaches to address these issues explore the exchange of knowledge between a human-in-the-loop and a learning system. This exchange, mediated through a teaching language, is often restricted to the specification of labels and constrains user expressiveness communicating other forms of knowledge to the system. We propose and assess an expressive teaching language for specifying object detectors which includes constructs such as concepts and relationships. From a formative study, we identified language building blocks and articulated design goals for creating interactive experiences in teaching object detection. We applied these goals through a design probe that highlighted further research questions and a set of design takeaways.