A U.S. Research Roadmap for Human Computation

The Web has made it possible to harness human cognition en masse to achieve new capabilities. Some of these successes are well known; for example Wikipedia has become the go-to place for basic information on all things; Duolingo engages millions of people in real-life translation of text, while simultaneously teaching them to speak foreign languages; and fold.it has enabled public-driven scientific discoveries by recasting complex biomedical challenges into popular online puzzle games. These and other early successes hint at the tremendous potential for future crowd-powered capabilities for the benefit of health, education, science, and society. In the process, a new field called Human Computation has emerged to better understand, replicate, and improve upon these successes through scientific research. Human Computation refers to the science that underlies online crowd-powered systems and was the topic of a recent visioning activity in which a representative cross-section of researchers, industry practitioners, visionaries, funding agency representatives, and policy makers came together to understand what makes crowd-powered systems successful. Teams of experts considered past, present, and future human computation systems to explore which kinds of crowd-powered systems have the greatest potential for societal impact and which kinds of research will best enable the efficient development of new crowd-powered systems to achieve this impact. This report summarize the products and findings of those activities as well as the unconventional process and activities employed by the workshop, which were informed by human computation research.

Learning to Predict the Wisdom of Crowds

The problem of "approximating the crowd" is that of estimating the crowd's majority opinion by querying only a subset of it. Algorithms that approximate the crowd can intelligently stretch a limited budget for a crowdsourcing task. We present an algorithm, "CrowdSense," that works in an online fashion to dynamically sample subsets of labelers based on an exploration/exploitation criterion. The algorithm produces a weighted combination of a subset of the labelers' votes that approximates the crowd's opinion.