CAREER: Symbolic Learning with Neural Language Models

Artificial intelligence (AI) systems today are very effective at learning statistical knowledge from large amounts of data. However, they are less effective at learning the forms of knowledge that we as humans might communicate in language to each other, such as the rules of a game, a food recipe, or the process of filling out your taxes. These forms of knowledge are symbolic, meaning that they can be represented as sentences in language, or, alternatively, as computer code. In this project the investigator will develop new AI methods for learning symbolic knowledge represented as computer code, combining ideas from statistics, large language models such as ChatGPT, and program synthesis (how to automatically generate computer software). The scientific impact of this research will be AI systems that learn more abstract forms of knowledge, from fewer examples, and which are more understandable to humans, because the systems will describe what they know in languages we can understand. This research will also involve student researchers, especially those from underrepresented groups. It will also inform new graduate and undergraduate classes, including the new Cornell undergraduate AI class, which serves around 150 students each semester. In more detail, this work addresses the problem of learning symbolic knowledge. Symbolic representations already form the cornerstone of automated planning, proof assistants, and other important applications, but the ability to learn symbolic knowledge is less mature compared to our ability to manually encode such knowledge. The work is organized around the observation that general-purpose programming languages like Python are very effective at representing certain kinds of symbolic knowledge, and also that pretrained neural language models are adept at generating such code. Based on these observations, the project adopts a framing that combines symbolic knowledge, Bayesian learning for uncertainty estimation, program synthesis, and neural language models for code generation and efficient probabilistic inference. The proposed work could ultimately benefit planning and model-based sequential decision-making, help us better understand human thinking and learning in computational terms, and take steps toward further automating software engineering. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.