Reading to Learn

Abstract

Traditional machine learning systems are trained on vast quantities of annotated data or experience. These systems often do not generalize to new, related problems that emerge after training, such as conversing about new topics or interacting with new environments. This thesis introduces Reading to Learn, a new class of algorithms that improve generalization by learning to read language specifications, without requiring any actual experience or labeled examples. This includes, for example, reading FAQ documents to learn to answer questions about new topics and reading manuals to learn to play new games. This thesis discusses new algorithms and data for Reading to Learn applied to a broad range of tasks, including policy learning in grounded environments and data synthesis for code generation, while also highlighting open challenges for this line of work. Ultimately, the goal of Reading to Learn is to democratize AI by making it accessible for low-resource problems where the practitioner cannot obtain annotated data at scale, but can instead write language specifications that models read to generalize.