- Machine learning from UW-Madison helps the New Yorker find the funny
- Decisions uses algorithms adapted from genomics research
- Process reduces editorial workload and helps engineers hone their technology
Since the famous New Yorker cartoon caption contest first launched as a weekly feature in 2005, cartoon editor Bob Mankoff and his team have personally sifted through more than 2 million entries in search of the funniest quips.
Reading and ranking thousands of entries each week is a daunting task, even for the most dedicated humorist. So when a data scientist from the University of Wisconsin-Madison proposed a technological solution, Mankoff was intrigued. The New Yorker’s signature brand of humor — sophisticated, erudite, playfully deadpan — is at the core of the 92-year-old magazine. Could an algorithm really be trusted to get the job done?
Rob Nowak, a professor of engineering who specializes in machine learning and statistical analysis, pitched Mankoff a computer program — essentially a crowdsourced survey — that allows users to vote on all the caption submissions, ranking them as either ‘funny,’ ‘unfunny,’ or ‘somewhat funny.’
As the votes come in, the captions getting the lowest marks are shown less frequently, and the ones that are considered funnier (and therefore more likely to be winners) are shown more often. Then, at the end of the voting period, the top three captions are published in the magazine and online and given another vote.
The algorithm, which has been powering the New Yorker contest since November 2015, sorts the captions based on the voting, giving each one a score and a probability that other voters would rank it similarly when compared to other captions.
“It does this, not perfectly, but better than any single person, even someone such as myself, could,” Mankoff says.
Nowak is the creator of NEXT, an open-source, cloud-based software system that allows users to quickly and easily develop applications for machine learning. Available free of charge on GitHub, NEXT technology is being used in about a dozen machine learning apps developed at UW-Madison as well as by the US Air Force, American Family Insurance, and Marshfield Clinic.
Machine learning uses computer algorithms to analyze and make predictions about data. Eventually, the computer is able to learn without being explicitly programmed.
NEXT software specializes in ‘active’ machine learning because the computer actively queries the user to get information that it needs. This is different from a traditional machine learning system, which involves users providing an annotated dataset and handing it over to the machine with the goal of the system learning a particular model.
“This system looks at machine learning from a different way,” Nowak notes. “it starts with no information, then the system decides what information it needs from people to form a good model. The goal is to reduce the burden on people.”
Nowak’s collaboration with the New Yorker is one of mutual benefit: The technology reduces the workload on Mankoff and his staff, and the data generated by the thousands of rankings helps Nowak test and perfect his technology. In fact, the algorithm used by the New Yorker was originally developed for an experiment on genetic diseases that aimed to determine a ranking of which specific genes were most important in expressing the disorder.
Employing machine learning to a field like genetics is just one example of how the technology can be applied across disciplines to greatly improve researchers’ capacity to analyze and understand experimental data. But perfecting the algorithms can be tricky.
“There’s always a little bit of a gap between the assumptions and the theoretical analysis, so experimental validation is super important,” Nowak admits.
But in many cases, running a scientific experiment is time consuming and expensive. That’s why the New Yorker caption contest is a perfect test case — it allows Nowak to run a new experiment every week.
“We’ve developed much better algorithms because we were able to keep running these experiments,” he says. “Algorithms that are working better for the New Yorker are also going to work better for these biological disease studies.”