Case Studies on the Interaction between Machine Learning and Language

Abstract

Designing machine intelligence is symbiotic with understanding human intelligence. We can use ideas from human cognition to inspire machine learning frameworks, and we can also use innovations in machine learning to better model human intelligence. Here, we consider how humans use language to inspire a new reinforcement learning framework, and we apply meta-learning algorithms to model how humans solve word analogies. In the first direction (Chapter 2), we note that humans use language to explain what they have learned, and that these explanations help other humans learn faster. We use this observation to inspire a teacher-student framework for reinforcement learning, whereby the teacher induces fast adaptation in the student by providing explanations. We show that our teacher-student framework enables student models to train significantly faster than training from scratch. In the second direction (Chapter 3), we use meta-learning over word embeddings to model human analogical reasoning, showing that our meta-learned model can match the performance of classical models such as the parallelogram model of analogy (king-man+woman=queen) and other heuristics while being trained on only 2406 samples. Our model also addresses some common critiques of using embeddings to model analogical reasoning, thereby demonstrating that vector space models still hold potential for understanding human cognition. Finally, we reflect on the commonalities between these two directions, noting that the symbiosis between machine and human intelligence is most fruitful when machines can or must solve problems in which the constraints of human intelligence apply (Chapter 4).