Speaker: Brenden Lake, New York University
Title: Three principles for building more human-like learning algorithms
There has been remarkable recent progress in AI and machine learning, yet many aspects of human learning still elude the best machine systems. I study computational problems that are easier for people than they are for machines. For example, people can learn a new concept from just one or a few examples, whereas machine learning algorithms typically need tens or hundreds of examples to reach similar levels of classification performance. People can also use their learned concepts in richer and more flexible ways than current machine systems -- for imagination, extrapolation, and explanation. Finally, people can learn by asking sophisticated and creative questions, whereas current active learning algorithms pose much simpler and more stereotyped queries. I will present my work on program induction as a cognitive model and potential solution to these computational challenges, with applications to learning handwritten characters, learning complex visual concepts (i.e., fractals), and asking questions in simple games. Across these case studies, my work suggests that three computational principles -- compositionality, causality, and learning-to-learn -- are important for understanding these human abilities and for building more human-like learning algorithms.
Brenden M. Lake is a Moore-Sloan Data Science Fellow at New York University. He received his Ph.D. in Cognitive Science from MIT in 2014, and his M.S. and B.S. in Symbolic Systems from Stanford University in 2009. He is a recipient of the Robert J. Glushko Prize for Outstanding Doctoral Dissertation in Cognitive Science, and his research was selected by Scientific American as one of the most important advances of 2016.
For additional information contact Steve Easterbrook