URSI project pushes bounds of artificial intelligence

For many, the idea of robots with the capacity to think like humans is for over-the-top, apocalyptic summer blockbusters starring Will Smith and has no place in everyday life in 2011. But while such individuals may believe Artificial Intelligence (AI) belongs solely in the realm of science fiction, Professor of Cognitive Science Ken Livingston and his team of student researchers share a different perspective. They have the notion that machines being able to process data in a humanistic way may be closer to reality than a layman might initially think.

For the past two summers, Livingston has worked with teams assembled through Vassar's Undergraduate Research Summer Institute (URSI) in order to explore the ways one can artificially create human-like intelligence. This past summer, he enlisted the help of Lily Pytel '13, Joshua Hawthorne-Madell '14 and Nina Vyedin '11.

"We are trying to replicate some of the features of primate cortex to see whether we can solve a certain set of problems in learning and intelligence," noted Livingston. "One of the big problems in robotics is building general purpose intelligence. We have lots of robots that do very specialized things. The Rumba will vacuum the floor, but if you ask it to make a cup of coffee, it can't do it. Watson can play a great game of Jeopardy!, but is completely incapable of doing anything else. So there have been a lot of tremendous successes in AI with robots that do narrowly focused tasks."

The project itself, formally titled Hierarchical Temporal Memory Solutions to the Robust Intelligence, specifically seeks to utilize our understanding of how the human brain processes complex tasks and to reproduce it in machines. This "bio-inspired" project, as Livingston refers to it, has physically manifested itself in an antenna-ed, Rumba-like research robot that processes information in a separate computer.

The link between robotics and neuroscience is not unexplored. Jeff Hawkins, creator of the Palm Pilot, was the first to found an institute for computer-based neuroscience research, known as the Redwood Center for Theoretical Based Neuroscience in Menlo Park, Calif. Hawkins is also credited as one of the first people to suggest that the human brain is physiologically hardwired to make predictions. When humans experience new situations, their brains store the information they gain as memory by building patterns of connections. For example, you learn at an early age that when you continue walking toward a solid object, such as a wall, without stopping, you will inevitably walk into it. Your brain then uses this pattern to predict outcomes when it encounters this situation again.

"The reason that we are so good at learning all kinds of different things is that the brain is a general purpose pattern-learning prediction maker. And the patterns it learns just depends on the experiences it has. The goal is to abstract away from the details of how the brain is doing things and figure out how to replicate that in the robot. Ultimately we want to get to the point where we can turn the robot loose in the world without having to program it," noted Livingston.

The project had a series of simple successes in which the robot could be presented a series of simple perceptual patterns and make predictions based on this information. Specifically, the robot would recognize a solid object in its path of vision and would know from past experiences not to collide with it. However, Livingston and his team were able to take their research one step further by obstructing its visual and audio input.

"It is sufficiently robust because our robot does something else that is historically difficult to get AIs to do; it learns the patterns even when the input is noisy. Like right now," referring to the squeaking of doors and the hum of Olmstead's industrial-strength air conditioner, "this is very noisy in the auditory channel. Humans can filter that junk out and just pick up the pattern of the language. For the robot, you can throw in up to 40 percent noise, or 40 percent junk, and it still picks up the pattern. However, it would take it longer to process the information."

The team also experimented with blurring and putting bars over the robot's field of vision. They were excited to find that they could occlude almost half of the image and the robot would still be able to recognize the object. However, Livingston quickly noted that the system is limited and any obstruction would result in the robot taking longer to process the information.

This summer was specifically devoted to optimizing the robot's processing speed and allowing it to better mimic the way in which a human cortex processes data. Among the improvements included programming the robot to recognize an object for what it is despite the angle at which it is viewed. This required the students to build a new network that is computationally simpler and faster than the one previously used.

Ultimately, Livingston hopes that such research could result in a robot that can learn to do complex tasks, such as cleaning your home. Think Rosie from The Jetsons rather than the mono-tasking Rumba. However, he is quick to note the potential drawbacks from such technology.

"I think there is no question that more general purpose kinds of systems are coming. The environments we encounter during the day, just in my lifetime, have gotten so much smarter. I mean, for crying out loud, my microwave is smarter compared to anything I knew as a kid. And as we come to depend on these technologies, there is a whole lot of our intelligence that we are offloading," said Livingston. He found these developments more troubling than exciting. "Take Google for example; who bothers to remember anything anymore when you can Google it?" he asked. "So I think we are changing the sort of creatures we are every time we offload some of that intelligence into the world. Almost nobody, for example, knows how to drive a standard transmission in a car anymore—that is all automatically computerized. Your brakes are antilock, with a fancy computer chip that knows the optimal way to break your car, so that's another skill you don't have to learn. There is a lot of information that we are shedding like that, and I think there are interesting questions about the point at which that becomes problematic."

Despite these potential social implications, Livingston notes that he has enjoyed working with his URSI students this summer. While robotics and AI have a long way to go before it is up to par with the expectations set for it by science fiction, their project, he feels, has been a valuable learning experience.