Woodpecker Test Gets DARPA's Attention


A computer program designed to better teach many people at once how to better distinguish between types of woodpeckers would certainly be a boon to the birding community, but the Defense Advanced Research Projects Agency (DARPA) is betting that it has wider-ranging benefits. 

DARPA, an agency of the U.S. Department of Defense in charge of developing new technologies for the U.S. military, has awarded its Young Faculty Award to Amin Karbasi, who developed the project. The highly competitive award is for $500,000 for two years, with a possible extension of another year and another $500,000.

Karbasi, assistant professor of electrical engineering & computer science, said the project aims to develop programs that learn how to better teach its users by observing them. It would be useful to DARPA for training many people at a time on a wide range of topics.

“It’s not like in a classroom with a teacher,” Karbasi said. “It’s one to one, it’s personalized. That is basically the dream of having massive online courses - to have a system that can teach many, many people and in a personalized manner.”

While computer training programs often take a “one size fits all” approach, Karbasi wants to develop a program that constantly monitors users’ progress with an automated teaching method that can adjust to users’ intents, abilities, interests, motivations, and learning styles.

“At DARPA, they need to know ‘Is this person OK, is this person not OK?’” Karbasi said. “It’s very hard for people to read and understand the cues, so they want an automated machine learning system that learns about humans. The challenge for us is: How can we train a machine such that, based on the information it gets over time, can understand what each cue means?”

Karbasi’s project, “Efficient Learning of Human Intent from Observations,” is based on preliminary results from one of his studies on training users to distinguish between three types of woodpeckers. Karbasi and his research team came up with an automated method that interacts with humans and uses examples to teach. Using Amazon’s Mechanical Turk platform, the program cues the user with pictures of the birds - they all look similar, but only one type of woodpecker is endangered. It then asks the user to identify the correct bird. 

“Based on the human answer, it then shows another example,” Karbasi said. “And the machine checks how you did, based on previous examples.” 

It then tries to give an example tailored to the user’s strengths and weakness, and designed to help the user learn faster.

“What we showed was that the machine can reason about the number of examples it has to show a human,” he said. “It is mostly about the interactions between the machine and the humans and about how the machines can teach humans by understanding their intent and where they’re making mistakes. DARPA is interested in these kinds of things because they want to train a lot of people at the same time who have different abilities and skills.”

Images (not those used in study) from the public domain.