Turnips or Bombs? Small talk with Professor Wand

January 29, 2016

If you run into Professor Mitchell Wand at a cocktail party, he might regale you with stories about a guy tucked away in “the bowels of the Pentagon,” writing a computer program.

“This guy’s job is to write a computer program,” Wand might say. “The job of that computer program is to take the images being sent back from drones over wherever it is we’re fighting at the present, to look at the images and to figure out whether that guy kneeling by the side of the road is planting turnips or planting a bomb.”

And right now, Wand’s work with the Probabilistic Programming for Advancing Machine Learning (PPAML) project is geared toward making that guy’s task simpler. The project, which has been in progress for over two years, is supported by the Defense Advanced Research Projects Agency (DARPA), a branch of the Department of Defense. Wand and Professor Olin Shivers are principal investigators on the project, which also includes other collaborators at Northeastern, at BAE Systems in Burlington and at Tufts University.

The crux of the project is the idea that advanced programming language technology can be employed to make the Pentagon guy’s job easier. Northeastern’s Racket programming system, developed by Professor Matthias Felleisen and other collaborators, is a major asset because it simplifies the process of creating new programming languages. A significant portion of the project is using mathematical models of programming languages to create a compiler for the new language, so the mathematical and theoretical expertise in the college is another important asset.

The new languages being developed for these applications are called “probabilistic programming languages.” These languages must incorporate classical programming language technology as well as probability and statistics. The mathematical techniques Wand and his collaborators rely on include Bayesian reasoning, a branch of probability theory devoted to figuring out how to incorporate new knowledge into an existing knowledge base.

“The guy in the basement is going to build a mathematical model of how to tell the difference” between a civilian and a militant, Wand says. “Our tools will help him express that mathematical model more easily and implement it as a computer program that is both easier for him to write and faster to run.”

Other applications for the project include robotics. If a robot with sensors and a camera is placed in an unfamiliar environment, it will set out to build a model of its surroundings. “With every step, it gets some information about the environment,” Wand explains. “It takes a step, it looks around again and says, ‘What’s new? What’s changed?’ It incorporates that new information into its model of the environment.”

About one-and-a-half years of funding remains for PPAML. At that point, DARPA will receive software – a system named Gamble – and several technical papers from the Northeastern team.

The biggest challenge Wand says he’s faced while working on PPAML is learning the mathematics and machine-learning technology necessary for the work. “Probabilistic programming and machine learning is a really new application area for us as programming language guys,” he says.