Departmental Seminar: Humanlike Few-Shot Learning using Bayesian Reasoning and Natural Language

Title: Humanlike Few-Shot Learning using Bayesian Reasoning and Natural Language

Abstract: Human inductive learning is rapid: From relatively few examples, we can learn a new rule in a game, or a new norm in a culture. Inductive learning is also broad: the space of learnable concepts is effectively unbounded, because simpler concepts can compose to build bigger ones. In this talk I propose an inductive learning model whose aim is to be more humanlike in that it is broader-coverage, while supporting learning that is rapid both in the number of examples and the amount of compute required. The model combines language models with Bayesian reasoning and neural code generation. It approaches human performance on two simple concept-learning domains, and gives a reasonably close fit to detailed human learning data. Together these results suggest an architecture for more human-like inductive learners, and a framework for aligning language-model inferences with human expectations by equipping language models with extra Bayesian machinery.