Is Code Enough? Stodden, Marinov Research Focuses On Providing Code Rather Than Generated Data With Research
Sharing and reusing research data is becoming increasingly common in the scientific world, allowing researchers to more easily build on the work of others as they seek new discoveries.
A new research project being conducted by Associate Professor Victoria Stodden of the School of Information Sciences and Illinois Computer Science Professor Darko Marinov aims to answer key questions about how researchers can reliably share the code used to generate their data rather than the more costly data itself.
“Our question was, when is it possible to save only the code that produced simulated data, and that’s all I need to save, and when do I also need to save the data?” Stodden said. “Simulation codes can produce massive amounts of data, for example petabytes of data. If I can rerun the code and regenerate the data, in theory I don’t even need to save the data. For what types of codes is that possible? That’s exactly the question we’re trying to answer.”
The National Science Foundation is funding the work by Stodden and Marinov, providing $300,000 over two years.
As Stodden, who is the lead investigator for the project and an Illinois Computer Science faculty affiliate, explains, the format for scholarly articles has changed little in decades. It provides only a small space to discuss how researchers derived their results.
But as computation has become more integral to research across virtually every scientific field, that format has become inadequate, she said.
“There’s such an amount of complexity – the computer can do X calculations per second. So how do you actually explain the increased complexity of computational research in words in a small section in a paper? It can be very, very difficult,” Stodden said.
Now some journals, she said, have begun to require researchers to publish their data and code along with their findings.
Stodden and Marinov, an expert on the testing and reliability of software, wondered whether providing the code alone could reliably allow the results of a given paper to be reproduced. And if it is the code that accompanies the published research, what kind of standard should it meet?
“If code is going to travel with this scholarly output, the community will need to come to some type agreement regarding code standards,” she said.
For their project, the two are focusing on physics research as an example because of its intensive computational needs.
In preliminary work using articles from the Journal of Computational Physics, Stodden and her group tried to replicate the computational results from 55 articles and were unable to reproduce any. After contacting the authors, Stodden says they came away with the impression that many believed reproducing their computational results would be straightforward, something they found not to be the case.
Eventually, Stodden and Marinov hope to determine whether and how code could be reliably substituted for data for a wide range of fields.
“We want to learn how to do better scientific software, software that is more reliable, and that researchers can trust more,” Stodden said. “These questions have come about not because the scientific community isn’t doing a good job; they came about because computation is so important, and increasingly so. We’re chasing fascinating opportunities here.”