Andrew Lampinen, Ph.D. student with Professor Jay McClelland, Department of Psychology, Stanford UniversityTitle: Meta-mapping: Toward human-like adaptability from deep learning
Abstract: Humans are able to flexibly adapt their behavior zero-shot (without data on the new task). For example, humans are told to try to lose at poker, they can do so on their first try, even if they have been trying to win in all their previous poker experience. By contrast, deep learning models generally cannot adapt flexibly. In this talk I will introduce meta-mappings, a computational framework for zero-shot adaptation. I will describe a parsimonious architecture that learns meta-mappings with the same networks used for basic tasks. I will demonstrate the success of this approach in card game, reinforcement learning, and visual concept settings, and will compare to alternative zero-shot approaches and human performance. Finally, I will show the utility of zero-shot adaptation as a starting point for later learning.We hope to see you all there!