Presentation Open Access
While algorithms and computing power get all the press, the special sauce behind many recent machine learning breakthroughs are meticulously labeled training data. Developing and maintaining these data sets as public goods is both an art and a science. In this talk I'll present a new set of best practices gleaned from interview with ~20 data set builders, maintainers, and funders. Topics include: encouraging collaboration between rival data teams; finding and addressing ethical issues with crowd labeling; launching competitions to spur data set use; and revenue generation models for sustainability.