Proceedings of the Annual Meeting of the Cognitive Science Society

People learn in fast and flexible ways that have not been emu-lated by machines. Once a person learns a new verb “dax,” heor she can effortlessly understand how to “dax twice,” “walkand dax,” or “dax vigorously.” There have been striking recentimprovements in machine learning for natural language pro-cessing, yet the best algorithms require vast amounts of experi-ence and struggle to generalize new concepts in compositionalways. To better understand these distinctively human abilities,we study the compositional skills of people through language-like instruction learning tasks. Our results show that peoplecan learn and use novel functional concepts from very fewexamples (few-shot learning), successfully applying familiarfunctions to novel inputs. People can also compose conceptsin complex ways that go beyond the provided demonstrations.Two additional experiments examined the assumptions and in-ductive biases that people make when solving these tasks, re-vealing three biases: mutual exclusivity, one-to-one mappings,and iconic concatenation. We discuss the implications for cog-nitive modeling and the potential for building machines withmore human-like language learning capabilities.