Human few-shot learning of compositional instructions

  • 2019-01-14 22:19:35
  • Brenden M. Lake, Tal Linzen, Marco Baroni
  • 13

Abstract

People learn in fast and flexible ways that have not been emulated bymachines. Once a person learns a new verb "dax," he or she can effortlesslyunderstand how to "dax twice," "walk and dax," or "dax vigorously." There havebeen striking recent improvements in machine learning for natural languageprocessing, yet the best algorithms require vast amounts of experience andstruggle to generalize new concepts in compositional ways. To better understandthese distinctively human abilities, we study the compositional skills ofpeople through language-like instruction learning tasks. Our results show thatpeople can learn and use novel functional concepts from very few examples(few-shot learning), successfully applying familiar functions to novel inputs.People can also compose concepts in complex ways that go beyond the provideddemonstrations. Two additional experiments examined the assumptions andinductive biases that people make when solving these tasks, revealing threebiases: mutual exclusivity, one-to-one mappings, and iconic concatenation. Wediscuss the implications for cognitive modeling and the potential for buildingmachines with more human-like language learning capabilities.

 

Quick Read (beta)

loading the full paper ...