Distill wants to be a sandbox for what a scientific paper can be
By Anna Nowogrodzki
Sometimes it’s hard to understand someone else’s research through a static scientific paper. Across countless universities and companies, at whiteboards and cafeteria tables, you’ll find scientists in animated conversations explaining their research to one another, asking questions, playing around with each other’s data: in short, interacting. Across the internet in recent years, people have extended these explanations to include interactive graphics and code.
Now a web-only machine-learning journal called Distill aims to provide a formal home for these interactive graphical explanations, which in recent years have expanded to blogs and other online fora.
“I think what they’re doing is really nice, and it’s kind of formalizing something that’s already happening informally,” says Jennifer Listgarten, a computational biologist at University of California at Berkeley. “People are posting blogs, or on GitHub repositories, IPython notebooks—things where you can actually really get a more intuitive feel than you can from just the traditional PDF.” Now the journal, which launched in March 2017, is making it possible to publish these communications directly by providing detailed design mentorship and custom web components that make building interactive papers easier than with raw HTML, CSS or JavaScript.
Distill’s editorial team, led by Google research scientists and co-founders Christopher Olah and Shan Carter, believes that interactive images and figures can often communicate better than static ones. Example figures have shown how a neural network decides whether an image is of a dog or a cat, and how momentum works. Readers can mouse over parts of a dog image to see trippy visualizations of what individual neurons in the network “see,” or drag slider bars adjusting step size and other variables to show how momentum creates its own oscillations.
By publishing these figures in an academic journal, Distill hopes to help researchers gain recognition for the work they do in creating them. “We saw over and over again people doing this kind of work and it not being able to help their academic careers,” says Olah.
The journal also aims to boost reproducibility by making backend code available so that readers can more quickly and easily replicate results. For Olah’s recent article about how machine learning “sees” and interprets photos, nearly 2,500 of the article’s 93,351 readers (2.6%) have interacted with the associated code notebook. That notebook allows readers to recreate the interactive graphics and reproduce the contributing analyses , says Olah, a feature that is available by simply clicking on links under each graphic that say “reproduce in a notebook.” “The paper is sort of under this permanent and ongoing review by the public,” Olah explains.
As technology editor Jeffrey Perkel recently wrote in Nature Toolbox, several other publications and publishers also support interactive graphics and code, including F1000Research, GigaScience, IEEE, and SPIE. But for Distill, dynamic graphics are not just supported but an area of primary focus. That approach is especially suited to machine learning, says Listgarten. “It’s very easy, especially in machine learning, for people to just … throw down a lot of equations and try to wow people with their technical prowess,” she says. “But at the end of the day, the best style of communication is often intuitive and pictorial.”
So far, only 10 papers have been published on Distill; all have been assigned DOIs. “We’re not publishing that many papers. We’re trying to create a prototype,” says Olah, of “a particular way of communicating scientific research. A lot of our papers are pushing in a new direction for what a paper can be.”
It’s not easy. Supporting dynamic figures involves intensive web development and design thinking, says Olah. Authors create the figures using tools such as the JavaScript D3.js library (for interactive figures) and Adobe Illustrator, Sketch or Inkscape (for static figures). Review of a paper includes not just traditional review elements, but also rigorous design feedback from Olah and Carter. The entire review process is documented in each paper’s GitHub repository, which is made public when the paper is published.
Olah, who has authored six papers in Distill, says the process is different from – and requires a greater time commitment than — traditional research publishing. “I’m investing so much more energy into it, sharing this way of thinking instead of just listing a bunch of results. It’s also a lot more work—I wouldn’t want to fool anyone. It’s a different kind of labor, and it’s as difficult as doing research.”
But it’s worth the effort, he adds. “I really like it when people are like, ‘Reading this was a pleasure.’ It feels like I got to make somebody happy.” And that’s not your typical scientific publishing experience.
Anna Nowogrodzki is a science writer based in Boston, Massachusetts.
Image credit: Goh, G. Why momentum really works. Distill (2017). https://doi.org/10.23915/distill.00006. Licensed under CC-BY 2.0.
Correction (8 May 2018): Jennifer Listgarten’s name was originally misspelled. Nature regrets the error.
Suggested posts
Software quality tests yield best practices
‘Manubot’ powers a crowdsourced ‘deep-learning’ review
eLife replaces commenting system with Hypothesis annotations
Recent comments on this blog
African astronomy and how one student broke into the field
From Doctorate to Data Science: A very short guide
Work/life balance: New definitions