Virginia Polytechnic Institute and State University researcher Arjun Chandrasekaran and colleagues report training a machine-learning algorithm to recognize and generate humorous scenes. They say the algorithm can accurately predict whether a scene is funny, even though it has no knowledge of the picture's social context.
The team asked workers on Amazon's Mechanical Turk service to produce funny scenes out of clip art, along with a short caption explaining the humor of the scene. They then were asked to create humorless scenes, and in this fashion the researchers compiled a database of 6,400 images that were 50 percent funny and 50 percent unfunny. Other Mechanical Turk workers were asked to rate how funny each scene was, and the resulting insights spurred the researchers to modify the 3,000 funny images in five ways to create a database of 15,000 unfunny counterparts of funny pictures. The machine-learning algorithm was trained to differentiate between funny and unfunny scenes by conducting two tasks--predicting the funniness of an image, and altering the funniness of an image by replacing an object in it. "The model learns...animate objects like humans and animals are more likely sources of humor compared to inanimate objects and therefore tends to replace these objects," the researchers note.
From "AI Algorithm Identifies Humorous Pictures"
Technology Review (01/08/16)
View Full Article
No entries found