NeST: The neuro-symbolic transpiler

作     者：Pfanschilling, Viktor Shindo, Hikaru Dhami, Devendra Singh Kersting, Kristian 

作者机构：Tech Univ Darmstadt Comp Sci Dept Hsch Str 1 D-64289 Darmstadt Germany Hessian Ctr Artificial Intellifence Hessian AI Landwehrstr 50a D-64293 Darmstadt Germany German Res Ctr Artificial Intelligence DFKI Landwehrstr 50a D-64293 Darmstadt Germany Eindhoven Univ Technol Dept Math & Comp Sci Room 7-142 NL-5612 AZ Eindhoven Netherlands 

出 版 物：《INTERNATIONAL JOURNAL OF APPROXIMATE REASONING》 (Int J Approximate Reasoning)

年 卷 期：2025年第179卷

核心收录：

学科分类：07[理学] 0701[理学-数学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 070101[理学-基础数学] 

基　　金：German Research Center for AI (DFKI) Hessian Ministry of Higher Education, Research and the Arts (HMWK) Federal Ministry for Economic Affairs and Climate Action (BMWK) AI lighthouse project "SPAICER" EU ICT-48 Network of AI Research Excellence Center "TAILOR" (EU) Collaboration Lab "AI in Construction" (AICO) Nexplore/HochTief EU Horizon project TANGO Federal Ministry of Education and Research (BMBF) Competence Center for AI and Labour [FKZ 02L19C150] HMWK cluster project "The Third Wave of AI" HMWK cluster project "The Adaptive Mind" Department of Mathematics and Computer Science and the Eindhoven Artificial Intelligence Systems Institute 01MK20015E 952215 

主　　题：Probabilistic programming Neuro-symbolic Large language models Tractable probabilistic models 

摘      要：Tractable Probabilistic Models such as Sum-Product Networks are a powerful category of models that offer a rich choice of fast probabilistic queries. However, they are limited in the distributions they can represent, e.g., they cannot define distributions using loops or recursion. To move towards more complex distributions, we introduce a novel neurosymbolic programming language, Sum Product Loop Language (SPLL), along with the Neuro-Symbolic Transpiler (NeST). SPLL aims to build inference code most closely resembling Tractable Probabilistic Models. NeST is the first neuro-symbolic transpiler-a compiler from one high-level language to another. It generates inference code from SPLL but natively supports other computing platforms, too. This way, SPLL can seamlessly interface with e.g. pretrained (neural) models in PyTorch or Julia. The result is a language that can run probabilistic inference on more generalized distributions, reason on neural network outputs, and provide gradients for training.