Our lab is investigating the basic questions of intelligence. What makes people smarter than machines? What ingredients enable the fast and flexible ways in which humans learn? Can we build machines that learn and think in more human-like ways?
In the last few years, there has been remarkable advances in machine learning and artificial intelligence. Computers have beaten Jeopardy champions, defeated Go masters, driven autonomous cars, and shattered records for object and speech recognition. Progress has been remarkable, and yet, the best example of intelligence is still natural intelligence. Human minds solve a diverse array of difficult computational problems every day: concept learning, object recognition, scene understanding, language acquisition, speech recognition, question asking, amongst many others. Machines also struggle to simulate other facets of human intelligence, including creativity, general purpose problem solving, and commonsense reasoning.
Our lab builds computational models of everyday cognitive abilities, focusing on computational problems that are easier for people than they are for machines. Almost by definition, these problems are interesting scientific pursuits for both cognitive science and machine learning. In cognitive science, if people outperform all existing algorithms on certain types of problems, we must learn how people solve them. In machine learning, these cognitive abilities are both important open problems as well as opportunities to reverse engineer the human solutions.
In this broad space of computational challenges, our work has touched on questions such as: How do people learn a new concept from just one or a few examples? How do people act creatively when designing new concepts? How do people learn qualitatively different forms of structure? How do people ask questions when searching for information? To get at these questions, we use computational modeling and behavioral experiments; we are also exploring neuroimaging and developmental studies. Across this range of questions and techniques, our work has revealed key cognitive ingredients that people use but are missing in contemporary machine learning. It has also led to new machine learning and data science techniques inspired by the cognitive solutions to these difficult computational problems.