Funds from this grant provide support for OpenMined, an online community of nearly 12,000 members from academia, industry, and government devoted to advancing privacy-preserving research methods in machine learning and AI development.   The OpenMined community is creating an ecosystem of advanced but accessible cryptographic tools designed to allow machine learning researchers to probe sensitive datasets without the need to copy, move or share any data.  Resources available on the OpenMined website (OpenMined.org) include a beginner’s guide, free classes and tutorials in a dozen languages, blogs and lectures from leading researchers in privacy-preserving research, and open-source coding repositories and projects on such topics as remote execution and federated learning, differential privacy, encrypted computation, and secure natural language processing.  Grant funds provide core operating support for the continued operation and expansion of the OpenMined community for a period of two years.

Funding decisions by the National Science Foundation and the National Institutes of Health are based on a process of peer review where independent subject-matter experts are invited to rate and rank potential research projects along a number of criteria, including the importance of the issue being examined, the strength of the experimental design, the likelihood of success, and the qualifications of the research team.  Projects that review well are funded.  Those that review poorly are not.  The budgets of just those two institutions total some $50 billion dollars per year—the lion’s share of basic research funding in the U.S.—so a lot rides on whether the peer review process is a good way to identify and select promising research projects.  This grant funds a series of projects led by Paula Stephan, Professor of Economics at Georgia State University and Research Associate at the National Bureau of Economic Research and Chiara Franzoni, Associate Professor at the School of Management at Milan Polytechnic University, to rigorously study peer review when used as a grant allocation process.  Stephan and Franzoni will partner with Novo Nordisk, a Danish foundation that has been funding research to combat diabetes since 1923 and that has kept meticulous records of its peer review and funding decisions.  This unique longitudinal dataset will allow the research team to compare the initial judgements of reviewers with the eventual successes of both funded and unfunded projects.  Second, Stephan and Franzoni have designed a series of experiments that will set up peer review panels with different decision structures and racial and gender compositions and give each panel the same set of research proposals to rate, giving the team some evidence of how structure and composition affect the verdicts of peer review committees.  Third, the team has assembled a small cohort of foundations that are willing to experiment with doing away with parts of the peer review altogether, introducing randomness at different stages of the selection process from a pool of proposals that meet certain base quality criteria.  These three separate initiatives will be supplemented alongside a host of qualitative and quantitative interviews that probe experts about the peer review process and their own judgements about its efficacy.