Giving Computers the Wisdom of People

July 29, 2014

by Suzanne Bouffard

Bias is usually thought to cause problems, but some kinds of bias can also help us solve problems. In fact, bias is an essential part of how we understand the world around us, says cognitive psychologist Tom Griffiths of the University of California at Berkeley. Every day, we are required to draw inferences, or educated guesses, about everything from what will happen if we turn a door knob to whether the foods we eat are safe. Our brains sometimes have little data from which to draw these conclusions, so they have to rely on our past experiences and our evolving beliefs about how the world works. Yet they are remarkably accurate. Griffiths’s work is illuminating why, and using that information to teach computers how to make inferences – one of the few areas in which they are less effective than humans.  

Griffiths studies inductive inference, or the process of drawing conclusions that go beyond the data available. “It’s very rare that we’re in a situation where we’re absolutely certain about the hypotheses that we’re making, so we use inductive reasoning all the time,” he points out. He cites examples from cognitive science including how people acquire language, how they understand how two objects will interact, and how they categorize things in their environments as edible or not.

How do we make these judgments? We use a range of prior knowledge, often subconsciously, including capacities we are born with and things we learn from other people or our own experience, Griffiths explains. In his research, he explores the idea that the way we draw inferences is related to the methods statisticians use to make sense of data. But unlike most statisticians, who have access to large amounts of data and are trying to form an unbiased opinion, we have to make inferences from small amounts of data, guided by our prior experience. This requires a form of statistics called Bayesian statistics.

Griffiths and his colleagues use behavioral experiments to test how well people’s inferences can be explained using Bayesian statistics. In particular, he has focused on understanding the biases that guide our inferences, and has developed a clever method for identifying those biases. It’s a laboratory procedure called iterated learning, and it works like the children’s game of telephone. Researchers observe a person as she learns new information, uses it to form a hypothesis, and passes on what she has learned to another person, and so on through a chain of people. Just as the message changes when passed from person to person in a game of telephone, the hypotheses people form change over time. And Griffiths and his colleagues have shown that these hypotheses change to be consistent with people’s biases, providing a clear picture of the effects of their prior knowledge on new inferences.

Griffiths uses the results of these studies to make mathematical models of how the human brain engages in inductive inference and then teaches computers how to “think” in a similar way. For example, using models based on human data, he has helped to develop a computer system that learns a new word, like “dog,” from just a few pictures and then applies it to other pictures of dogs. Although people make these judgments all the time, this kind of reasoning has historically been a big challenge for computers. Griffiths hopes to change that, and thinks that computers will ultimately be able to solve important problems like making medical diagnoses and scientific discoveries.

“We can try to give computers the wisdom of people,” he believes, and because they are capable of operating at much higher speeds and larger scale than people, that means the potential for big leaps in scientific discovery. “We’re expanding the set of things computers can do,” he explains, “which expands the set of things humans can do.”  




tom griffiths 174X232.jpgTom Griffiths will be honored with the Federation of Associations in Behavioral & Brain Sciences (FABBS) Foundation Early Career Impact Award during the 2014 annual meeting of the Cognitive Science Society. 






To learn more about this area of science: