Exploring empty spaces: increased human data in the loop

Data increase is crucial to make automatic learning models more robust and safe. However, increasing the data can be a challenge, since it requires generating various data points to rigorously evaluate the behavior of the model in edge cases and mitigate possible damage. Creating high quality increases that cover these “unknown unknowns” is an intensive task in time and creativity. In this work, we present to broad, an interactive tool to help professionals navigate “unknown unknown” in unstructured text data sets and improve data diversity by systematically identifying empty data spaces to explore. Wide includes three human data increase techniques in the circuit: increase with concepts, increase by interpolation and increase with a large language model. In a user study with 18 Teamers Red Professional, we demonstrate the usefulness of our increased methods to help generate high quality, diverse and relevant models indications. We found that wide allowed the red teams to increase the data quickly and creatively, highlighting the transformative potential of the interactive increase workflows.