Learning to Reason: Decomposition, Planning, and Quantitative Reasoning
Abstract
Approved for Public ReleaseThe recent success of pretrained language models (LMs) has brought a lot of excitement to the Natural Language Processing (NLP) community and the AI community in general; the media, pushed by some companies, is discussing again how close AI is to human performance and how close we are to "general" AI. While this success has led to a paradigm shift in some parts of NLP research, all scientific conferences are hosting panels discussing the limitation of transformers, the key building block of all current models, the limitations of our current training procedures, and its consequences to the problem the research community is addressing, and those that are left behind. The research community is busy looking for the coin under the lamppost, going for easy wins, and neglecting to study some of the fundamental problems that NLP and AI need to address to make real progress. The thesis pursued in this proposal is that making decisions that depend on natural language understanding requires reasoning abilities, that depend on multiple, interdependent, models (sometimes it is useful to think about it as "symbols"). This cannot be accomplished by "evaluating" a single model nor can we train directly to accomplish it. Rather, we argue that at the heart these decisions is a planning process that determines what modules are relevant and what knowledge needs to be accessed in order to support the decision. To accomplish it, our intelligent agents need to decompose (to exploit compositionality), compose (to build on independently learned models), and plan.The proposal first discusses the types of reasoning LLMs cannot do and suggests that the common obstacle among these is the need to identify intermediate representation and perform reasoning on it, rather than on the text itself. Examples of these tasks include reasoning problems that require aggregation (how many US presidents had more than 2 daughters?), those that require the use ofcertain logical operators, and those the require planning. Our primary scientific goal is to addresses some of the key scientific challenges that prevent us from moving beyond the narrow training paradigm of current LLMs facilitates and develop ways to support a range of reasoning tasks that are essential to support intelligent applications. After discussing the limitations of current paradigms, we propose to study two fundamental and complementary themes. On the Inference side, we propose a way to identify the propositions (variables) that are involved in the reasoning process. This is a process we call "decomposition" and it essentially identifies the intermediate representation that is needed in order to support reasoning tasks. This process relies on a knowledge acquisition step that depends on the identification and exploitation of incidental supervision signals. To complement this, on the Learning side, we suggest studying the theory of learning with indirect supervision - this is an important complement since in almost all cases we cannot supervise the intermediate representation; we can only supervise a function of the propositions represented there. The PI is a leading researcher in Artificial Intelligence, with expertise and major contributions in natural language processing, machine learning, and knowledge representation and reasoning.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312364
Entities
People
- Dan Roth
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania