The perky MIT professor rides to work on a bicycle that neatly folds into a briefcase-sized rectangle and spends her days trying to make machines that can learn. As associate director of MIT's Artificial Intelligence Laboratory, she is surrounded by quirky creations, from tiny fish-like robots made partly from frog tissue to robots that look like humans and, one day, may think like them, too. But Kaelbling leads another revolution in her spare time. As editor of the upstart Journal of Machine Learning Research, Kaelbling offers some of the latest debates and developments in artificial intelligence to anyone with access to the Internet - free. By contrast, the scholarly journals that largely set the world's science agenda sometimes charge more than $1,000 a year for a subscription.
Kaelbling would be just another quixotic idealist with her unpaid staff - the journal "doesn't even have a bank account right now," she acknowledges - except for one thing. Thirty-nine board members of the leading artificial intelligence journal in her field, Machine Learning, announced their resignations last month to join her crusade.
Kaelbling and her supporters argue that putting all top current research online could well inspire crucial insights from people who wouldn't otherwise have access; a statistician in Mongolia may spur the next breakthrough, after all.
"The only thing we care about in the world is that people read our work," Kaelbling said.
The rivalry between Kaelbling's journal and Machine Learning - which costs $1,050 a year for institutions and $120 for individuals - is part of a much broader debate about how much scientific information should be free. For instance, the federal government's Human Genome Project and private Celera Genomics have locked horns repeatedly over Celera's plan to withhold some key information from the human genetic blueprint that could be sold to pharmaceutical companies looking for potential new drugs.
More closely parallel to Kaelbling's work, the National Institutes of Health have created a database called PubMed Central intended to allow anyone in the world to freely search and retrieve the full text of any published scientific article, with archives extending back for decades. Several months ago, a coalition of about 30,000 scientists, led partly by former NIH director and Nobel laureate Harold Varmus, pledged to boycott any journals that didn't meet PubMed Central's standards for freely distributing data. The coalition backed off a little in early September, but the conflict still looms.
The free, Internet-based Journal of Machine Learning Research challenges an elaborate system of disseminating scientific information that touches everything from what appears on the nightly news to which researchers become stars. A few well-known journals such as Science, Nature, and the New England Journal of Medicine considerably influence scientific discussion and have a near hammerlock on determining what science appears in the mainstream press. A slew of others - including Machine Learning - shape their respective subject areas, helping determine who gets tenure, where grants go, and how their fields move forward.
Kaelbling and others note that the scientific community differs in several ways from journalism, where support for free online access to magazines and newspapers has withered. For one, most scientific authors don't get paid for publishing, even in the journals with costly subscription prices. Instead, they receive their funding from universities, corporations, or government grants, and publish mainly for prestige and to advance their fields. Kaelbling earns her salary from the Massachusetts Institute of Technology and part of her job description requires her to offer public service to the community, such as editing the Journal of Machine Learning Research.
Secondly, progress in any scientific field relies to a huge extent on the amount of available information. More available information equals more and better science. Some of Kaelbling's colleagues, for example, want to design robots that actually think like humans: Rodney Brooks, the director of Kaelbling's laboratory, helped inspire Steven Spielberg's vision for the movie "A.I." That enormously complicated task would be helped with input from as many different scientists as possible. Putting everything online seems like the obvious thing to do, Kaelbling argued.
Still, numerous scientists and publishers fear that moving away from printed, subscription-based journals could derail standards and practices that have worked well for years. Others fear that such a change could shutter prestigious and important journals.
In the machine-learning community, swords have already crossed. In their letter resigning from Machine Learning, the rebellious members, who represented about two-thirds of the board, wrote: "Journals should principally serve the needs of the intellectual community, in particular by providing the immediate and universal access to journal articles that modern technology supports, and doing so at a cost that excludes no one."
In the past, scholars in the field researched and wrote their articles, submitted them to Machine Learning, and then waited up to a year to see them in print. Probably most aggravating to the authors, Machine Learning's owners, Netherlands-based Kluwer Academic Publishers, retained complete copyright control. Authors couldn't even publish their articles on personal Web pages - a fairly restrictive policy for the industry that the company changed after the mass resignation.
With the Journal of Machine Learning Research, authors just e-mail their pieces to Kaelbling, who then forwards them to assorted editors. These volunteers, generally prestigious researchers in whatever particular sub-field the article covers, then decide whether to accept or reject the articles. If accepted, the articles appear online immediately and the authors retain full copyrights.
"Mostly, it's just a bunch of work," said Kaelbling, before noting that she still spends vastly more time with her MIT students and Erik the Red, a robot resembling R2D2 that she is trying to teach to see and navigate through hallways.
Despite Kaelbling's optimism, though, other scientists argue that there are holes in her boat. Robert Holte, the editor of Machine Learning, supports the new journal and suggests that both his journal and Kaelbling's can exist harmoniously. But, he added, "What [the Journal of Machine Learning Research] doesn't have right now is a history. You can have the most famous people on your editorial board that you like. But until a journal has a well-established track record proving its ability to attract a large number of high quality, highly-cited papers, it cannot claim to be the community's flagship journal."
Tenure committees, for example, know that being published in Machine Learning means you've accomplished something significant. Until it earns a reputation, Kaelbling's journal could just be two crackpots in a barn with a cable modem.
In addition, advocates of print journals argue that online journals may not hold to the same quality standards, becoming, in a sense, the scientific equivalents of the Drudge Report.
"Paper journals have a strict limit on the number of papers they can publish. With an online journal, there's always a temptation to accept rather than reject," said Cornell professor Shai Ben-David, one member of the editorial board of Machine Learning who chose not to resign.
Kaelbling acknowledged that her journal hasn't earned prestige yet, but she insists that it will maintain strict standards. "It's the same people doing the same work," she said, noting that the editorial board of the Journal of Machine Learning Research is made up of many people who used to work for Machine Learning. Kaelbling is one of the most highly respected scientists in her field and, before resigning last year, she herself reviewed papers for her print rival.
Kaelbling faces another potential problem in that no major for-profit publisher supports and promotes her journal. Kluwer Academic Publishing, the world's second largest scientific publisher with 731 journals under its umbrella, stated that it supports Machine Learning by providing services that include promotion, copy-editing, distribution and representation of the journal at conferences.
Still, the nonprofit MIT Press does offer Kaelbling's journal its support, publishing and promoting quarterly bound editions of the articles that have appeared on the journal's Web site. "I don't think they are any less effective than Kluwer," Kaelbling said.
MIT Press does publish and promote quarterly bound editions of the articles that have appeared on Kaelbling's Web site, but has garnered fewer than 100 subscriptions so far. That doesn't phase Kaelbling, either. "Everyone's going to have to do this," she said.
Even some scientists closely attached to old print publications agree. According to Thomas Dietterich, a former editor of Machine Learning, and one of the recent defectors to the new journal: "I am emotionally attached to Machine Learning. I have every issue from the start to now. But things change. In the computer business, we are used to technology turning things upside down."