Leslie Pack Kaelbling’s interest in artificial intelligence began during her senior year of high school, when she discovered the book Gödel Escher Bach: An Eternal Golden Braid. Published in 1979, the philosophical work ponders the mystery of how the mass of cells comprising the human brain gives rise to intelligence and consciousness, and whether computers can ever be made to mimic that phenomenon.
“The book got me excited,” says Kaelbling, who is now the Ellen Swallow Richards Professor of Electrical Engineering and Computer Science and a faculty leader of the recently launched MIT Intelligence Initiative, which brings cognitive scientists and computer scientists together to explore the nature of intelligence and create machines with more human-like intelligence. The project could propel Kaelbling toward her dream of developing a robot capable of learning and decision-making in many situations.
“Because humans can learn and plan, I firmly believe that I can make a machine do that, too,” she says.
Kaelbling’s work focuses on a branch of artificial intelligence (AI) known as machine learning. Her research on computer systems that adapt to a complex, changing environment has contributed to innovations in mobile robotics, as well as programs that help airplanes avoid collisions, provide intelligent assistance on the computer desktop, and help drivers improve fuel economy.
Yet, Kaelbling and her colleagues have faced a metaphoric brick wall when trying to design machines that can perform certain tasks that come naturally to humans but are deceptively complex. For example, she is working on a “seeing” robotic arm that learns to grasp a bottle, first by determining its three-dimensional structure through video from a webcam, then by repeatedly sensing the bottle’s physical properties through touch.
“Vision remains incredibly difficult,” Kaelbling says in describing the challenges of training machines to accomplish something that a baby can do instinctively. “Humans are expert at vision in a way that’s hard to understand. It’s so deeply wired.” She confronts this challenge by approximating human vision through video, and augmenting that capability with a robust grasping system. “It’s a question of understanding not just vision by itself, but what it needs to deliver to the downstream tasks,” she explains.
Such strategies, along with machine-learning approaches that rely more on experience and probability than on top-down logic, are needed to create a robust and flexible robot, Kaelbling believes. She envisions “a sidekick that would watch you do things, infer what your strategy is, and assist you accordingly” in myriad tasks, such as tidying your living room or anticipating conflicts in your calendar six months out. Such a machine must combine many intelligent capabilities — vision, dexterity, mobility, maneuverability, language comprehension — in addition to the ability to learn, adapt to new situations, and make decisions.
Kaelbling says that it is useful to learn from neuroscientists, who want to discover exactly how the human brain is structured and how it functions. But, she says, “from an engineering perspective, it might be very difficult to understand and recreate in computers what millions of years of evolution produced in humans.” Instead, she employs a pragmatic strategy, creating machine intelligence through techniques from computer science, statistics, and mathematics. “A direct engineering approach might create a solution that’s very different from the evolutionary one,” she says, “but it is likely to bear fruit sooner.”