Machine Learning, Understanding, and Interaction

High-performance machine learning models generally suffer from a lack of transparency that calls into question the ability of humans to understand how they work. Through the study of CoMo, an interactive learning system that allows users to associate sounds and gestures, we show that a form of model understanding can emerge from user-model interactions. This interactive approach offers an alternative to explanation- and interpretation-based epistemologies. Although data- and computation-intensive deep learning systems are not well suited to this form of interactivity, shallow learning models offer underappreciated epistemological advantages in contexts where machine learning can be made interactive.