AI holds the potential to assist doctors find early markers of disease and policymakers to avoid decisions that result in war. But a growing body of evidence has revealed deep flaws in how machine learning is utilized in science, an issue that has swept through dozens of fields and implicated 1000’s of erroneous papers.
Now an interdisciplinary team of 19 researchers, led by Princeton University computer scientists Arvind Narayanan and Sayash Kapoor, has published guidelines for the responsible use of machine learning in science.
“Once we graduate from traditional statistical methods to machine learning methods, there are a vastly greater number of how to shoot oneself within the foot,” said Narayanan, director of Princeton’s Center for Information Technology Policy and a professor of computer science. “If we do not have an intervention to enhance our scientific standards and reporting standards with regards to machine learning-based science, we risk not only one discipline but many alternative scientific disciplines rediscovering these crises one after one other.”
The authors say their work is an effort to stamp out this smoldering crisis of credibility that threatens to engulf nearly every corner of the research enterprise. A paper detailing their guidelines appeared May 1 within the journal Science Advances.
Because machine learning has been adopted across virtually every scientific discipline, with no universal standards safeguarding the integrity of those methods, Narayanan said the present crisis, which he calls the reproducibility crisis, could develop into much more serious than the replication crisis that emerged in social psychology greater than a decade ago.
The excellent news is that a straightforward set of best practices can assist resolve this newer crisis before it gets out of hand, in accordance with the authors, who come from computer science, mathematics, social science and health research.
“This can be a systematic problem with systematic solutions,” said Kapoor, a graduate student who works with Narayanan and who organized the hassle to supply the brand new consensus-based checklist.
The checklist focuses on ensuring the integrity of research that uses machine learning. Science is dependent upon the power to independently reproduce results and validate claims. Otherwise, latest work can’t be reliably built atop old work, and all the enterprise collapses. While other researchers have developed checklists that apply to discipline-specific problems, notably in medicine, the brand new guidelines start with the underlying methods and apply them to any quantitative discipline.
One in every of the most important takeaways is transparency. The checklist calls on researchers to supply detailed descriptions of every machine learning model, including the code, the info used to coach and test the model, the hardware specifications used to supply the outcomes, the experimental design, the project’s goals and any limitations of the study’s findings. The standards are flexible enough to accommodate a big selection of nuance, including private datasets and complicated hardware configurations, in accordance with the authors.
While the increased rigor of those latest standards might slow the publication of any given study, the authors imagine wide adoption of those standards would increase the general rate of discovery and innovation, potentially by lots.
“What we ultimately care about is the pace of scientific progress,” said sociologist Emily Cantrell, considered one of the lead authors, who’s pursuing her Ph.D. at Princeton. “By ensuring the papers that get published are of top of the range and that they seem to be a solid base for future papers to construct on, that potentially then quickens the pace of scientific progress. Specializing in scientific progress itself and not only getting papers out the door is de facto where our emphasis must be.”
Kapoor concurred. The errors hurt. “On the collective level, it’s just a significant time sink,” he said. That point costs money. And that cash, once wasted, could have catastrophic downstream effects, limiting the sorts of science that attract funding and investment, tanking ventures which might be inadvertently built on faulty science, and discouraging countless numbers of young researchers.
In working toward a consensus about what must be included in the rules, the authors said they aimed to strike a balance: easy enough to be widely adopted, comprehensive enough to catch as many common mistakes as possible.
They are saying researchers could adopt the standards to enhance their very own work; peer reviewers could use the checklist to evaluate papers; and journals could adopt the standards as a requirement for publication.
“The scientific literature, especially in applied machine learning research, is filled with avoidable errors,” Narayanan said. “And we would like to assist people. We would like to maintain honest people honest.”
