sLTN：结构化逻辑张量网络

文章背景与核心概要

逻辑张量网络（Logic Tensor Networks, LTN）作为神经符号AI领域的重要框架，长期以来主要用于处理扁平化（flat）的数据集合。然而，面对现实世界中复杂的时序、序列以及图结构数据时，传统LTN在表达能力上往往捉襟见肘。为了解决这一痛点，由 Davide Rinaldi 和 Luciano Serafini 推出的 sLTN（Structural Logic Tensor Networks）引入了作为“一等公民”（first-class citizens）的“结构维度”。

该论文的核心技术贡献在于：通过将张量轴（如时间步、序列位置或图节点）作为逻辑语言中的显式组成部分，sLTN 实现了在逻辑层面对复杂关系、时序和序列约束的直接表达。同时，该框架保持了良好的向后兼容性——在没有结构维度的特例下，能够完美退化并恢复为传统的 LTN 语义。作者团队还开源了基于 PyTorch 的实现库（包含声明式签名和公式解析系统），极大地便利了研究人员在神经架构中开展结构化推理。

sLTN：结构化逻辑张量网络

sLTN: Structural Logic Tensor Networks

作者： Davide رinaldi（Davide Rinaldi）, Luciano Serafini
日期： 2026年8月11日
学科： 人工智能 (cs.AI)
arXiv ID: 2608.11136

摘要

Summary sLTN (Structural Logic Tensor Networks) is an advanced neurosymbolic framework that extends the traditional Logic Tensor Network (LTN) paradigm. While standard LTNs are designed for flat data collections, sLTN introduces "structural dimensions" as first-class citizens within the logical language.

By treating tensor axes—such as temporal steps, sequence positions, or graph nodes—as explicit components, sLTN allows for the direct expression of complex relational, sequential, and temporal constraints at the logical level. The framework maintains backward compatibility, recovering original LTN semantics when structural dimensions are absent. The authors provide a PyTorch-based implementation, including a declarative signature and formula parsing system, to facilitate structural reasoning in neural architectures.

核心特性

Key Features * Structural Dimensions: Integrates domain-specific organization (time, sequence, graph connectivity) directly into the tensor-based logical language. * Enhanced Expressivity: Enables explicit quantification and relational constraints across structured data. * Formal Semantics: Provides a rigorous fuzzy tensor semantics for structural logic. * Implementation: Accompanied by the sltn library, supporting declarative formula definition and automated parsing.

访问与资源

Access & Resources * View PDF * Source Code (GitHub) * HTML Version (Experimental)

许可协议

License View License (CC BY 4.0)