Big Questions: Can Robots Learn Common Sense?

—Katharine Reece MFA ’12

If you place your phone on a table, does it fall through to the floor? Of course not. But a robot doesn’t know that. Computers can now beat humans at chess, solve the most complex math equations, and anticipate the kind of music you might like, but they have very little common sense. “We take for granted how complicated a task the brain has,” says Jim Marshall (computer science), SLC’s robotics expert. Training a computer to mimic the human brain—to not only respond to stimuli, but also recognize patterns and learn from experience—requires wrestling with the as-yet-unanswered question: What is intelligence?

“If you can write programs that simulate a bunch of simple elements connected together, you can teach a computer program to learn from experience.” —Jim Marshall (computer science)

Rewind to the late ’70s, before the term “personal computer” became commonplace. Marshall was a teenager in northern Virginia, saving up money from a paper route to buy his first computer, a RadioShack TRS80. The idea of an intelligent computer had captivated him since Arthur C. Clarke’s novel 2001: A Space Odyssey alarmed the world with its portrayal of HAL, the self-aware, homicidal computer. What would prevent computers from surpassing human intelligence and ultimately leaving us in the dust?

The intelligence of computers has indeed increased exponentially in the last few decades, but Marshall isn’t worried about a robot takeover. While their computational power may be staggering, they still can’t do most tasks humans execute with ease, such as walking around without bumping into things, responding to human emotion, or understanding the implications of gravity.

So where does common sense come from? There’s no pat answer, but it has to do with the ability to learn. In “Bio-Inspired Artificial Intelligence,” Marshall teaches students to program terrier-sized Sony dog robots, a humanoid named Marvin, and a machine with wheels that resembles Spirit, the robot that explored Mars in 2003. Programming these robots to take in sensory information and respond to it is only the first step. Then comes ­Marshall’s true passion: developmental robotics, in which you teach the robots to learn by creating programs modeled on the configuration of the brain. “If you can write programs that simulate a bunch of simple elements connected together,” Marshall says, “you can teach a computer program to learn from experience.”

One might argue that the limitations of computers are actually the limitations of humans; we’re constrained by the mysteries of human experience. For Marshall, that’s where the beauty of this evolving field shines brightest. He wonders if perhaps intelligence is actually simpler than we think. “We’re all made out of, what, 92 elements?” he asks. “Essentially, everything you see about the world involves patterns of very simple components, but the world is a complex, amazing place nevertheless. Maybe intelligence is similar to that. Maybe there are fundamental principles that are inherently simple that we don’t understand yet.”