Humor 101: A Virginia Tech research team tackles the algorithm of humor
November 9, 2016
Computers being trained for humor quickly learned that humans and animals tended to make scenes humorous, so the computer usually chose to replace those types of objects when trying to make a scene less funny.
Are you sad because no one laughs at your jokes Take heart. The day is coming when a robot will! Or, then again, maybe it won'teven the computer thinks you're corny.
Members of a Virginia Tech research team are harbingers of this humor apocalypse. Arjun Chandrasekaran, a Ph.D. student in computer engineering, and Ashwin Kalyan, a visiting student from the National Institute of Technology Karnataka in India who will be starting as a Ph.D. student in fall 2016, are the lead authors on a paper entitled "We are humor beings: Understanding and predicting visual humor," which has gotten press in magazines like the MIT Technology Review and Newsweek.
The project investigates how a computer could learn to recognize and replicate humor in visual scenes. This is a collaboration between ECE's Computer Vision Lab and the Machine Learning and Perception Group, led by assistant professors Devi Parikh and Dhruv Batra, respectively. Parikh and Batra are co-authors of this study, along with Mohit Bansal from Toyota Technological Institute, and Larry Zitnick from Facebook AI Research.
"Humor is such an important part of the human experience," said Parikh. "But, surprisingly, we still don't have a grasp on why one scene is funny while another is notespecially when it comes to visual humor."
Equipping a machine with a sense of humor may be a long way off, but the collaboration is approaching this task by presenting the computer with examples of scenes. They are accumulating a database of images that run the gamut from mundane to hilarious.
The team employs workers from Amazon Mechanical Turk, an online marketplace for work, to create their own funny scenes from clip art and include a short description of why they think the scenes are funny.
The team calibrated the database by asking workers to rate the funniness of each scene and found that, for the most part, everyone agreed on which images were actually funny. They also found that an unfunny scene could take on a comedic edge when the objects were switched aroundsubstituting a raccoon for a man, for instance, or a piece of cheese for a cat.
To train an algorithm to spot the difference between funny and unfunny images, the team gave the machine two tasks: judge the funniness of a scene, and then alter the funniness of a scene by replacing an object within it.
The machine's algorithm beat the baseline, which means the computer had a general notion of what was supposed to be funnycertainly better than a random guess.
The model learned that, in general, animate objects like humans and animals are more likely to be sources of humor than inanimate objects, and therefore tended to replace these objects, the team reported.