Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 45

On the Representational Geometry of Dynamic Programs 动态程序的表示几何学

Standard neural architectures struggle to generalize to longer inputs for dynamic programming (DP) targets, and this paper investigates the geometric root causes Every finite min-plus DP can be equivalently described as a shortest path on a DAG, a tropical polynomial, or an extended Newton polyhedron — and these three descriptions form isomorphic semirings at both formal polynomial and computed function levels The paper proves that the two native dimension-reduction operations in the semiring (s 标准神经网络难以泛化到更长输入的动态规划(DP)问题,本文从几何角度揭示其根本原因 证明每个有限min-plus DP等价于DAG上的最短路径,且可表示为热带多项式,其扩展牛顿多面体编码决策边界 建立图、多项式、多面体三种描述的半环同构关系,刻画所有结构冗余 证明长度泛化存在结构性障碍:维度约简操作既不单射也不封闭,串并联组合无法构造所有DAG拓扑

58
Hot 热度
72
Quality 质量
63
Impact 影响力

Analysis 深度分析

TL;DR

  • Standard neural architectures struggle to generalize to longer inputs for dynamic programming (DP) targets, and this paper investigates the geometric root causes
  • Every finite min-plus DP can be equivalently described as a shortest path on a DAG, a tropical polynomial, or an extended Newton polyhedron — and these three descriptions form isomorphic semirings at both formal polynomial and computed function levels
  • The paper proves that the two native dimension-reduction operations in the semiring (setting variables to identities) are neither injective nor always closed within DP, explaining a fundamental generalization barrier
  • Series and parallel composition of DAGs fail to construct all possible DAG topologies from smaller sub-DAGs, and even terminal-only operations cannot capture all DP compositions
  • These structural negatives provide a rigorous geometric explanation for why length generalization in DP tasks remains an open challenge for neural networks

Why It Matters

This work bridges discrete mathematics, algebraic geometry, and machine learning by providing a formal geometric framework to understand a persistent practical problem: why neural networks trained on DP tasks fail to generalize beyond their training sequence lengths. For AI practitioners building models on algorithmic reasoning tasks, this paper offers theoretical grounding for observed failures and points toward structural limitations that architectural changes alone may not overcome.

Technical Details

  • Three isomorphic descriptions: The paper establishes that finite min-plus DP problems admit three equivalent formulations — graph-theoretic (shortest path on a DAG), algebraic (tropical polynomial), and geometric (extended Newton polyhedron encoding decision boundaries) — with isomorphic semiring structures at both the formal polynomial level and the computed function level
  • Structural redundancy characterization: Operations connecting the two semiring levels are shown to characterize all structural redundancies in DP representations, providing a complete algebraic taxonomy of DP equivalences
  • Negative generalization results: Two key structural negatives are proven — (1) dimension reduction via identity substitution is neither injective nor closed within the DP semiring, and (2) series/parallel composition and terminal-only operations are insufficient to construct all DAG topologies or capture all DP compositions
  • Decision boundary analysis: The extended Newton polyhedron encodes the decision boundary of which path wins in a DP, and the paper frames the length-generalization question as whether the decision boundary at length T determines the boundary at length T+1

Industry Insight

  • Researchers developing neural solvers for algorithmic reasoning should treat length generalization in DP as a fundamental structural limitation rather than a mere training or architectural deficiency; model design should account for the proven incompleteness of series/parallel composition in capturing DP structure
  • The isomorphic semiring framework could inform the design of inductive biases or structured neural architectures that respect the algebraic structure of DP problems, potentially improving generalization by aligning model capacity with the true complexity class of the target
  • This theoretical groundwork suggests that hybrid approaches combining neural networks with explicit symbolic or algorithmic components may be necessary for robust DP generalization, as pure end-to-end learning faces provable geometric barriers

TL;DR

  • 标准神经网络难以泛化到更长输入的动态规划(DP)问题,本文从几何角度揭示其根本原因
  • 证明每个有限min-plus DP等价于DAG上的最短路径,且可表示为热带多项式,其扩展牛顿多面体编码决策边界
  • 建立图、多项式、多面体三种描述的半环同构关系,刻画所有结构冗余
  • 证明长度泛化存在结构性障碍:维度约简操作既不单射也不封闭,串并联组合无法构造所有DAG拓扑

为什么值得看

本文首次从代数几何视角系统分析动态规划问题的表示学习困难,为理解神经网络在组合优化任务上的泛化瓶颈提供了理论框架。研究结果对设计更有效的结构感知神经网络架构具有重要指导意义。

技术解析

  • 核心等价性:证明有限min-plus DP、DAG最短路径、热带多项式三者之间存在严格的半环同构关系,在形式多项式和计算函数两个层面均成立
  • 决策边界几何:扩展牛顿多面体完整编码了"哪条路径获胜"的决策边界,为分析泛化能力提供了几何工具
  • 泛化障碍证明:证明两种维度约简操作(变量设为单位元)既不单射也不在DP内封闭,从代数结构上解释了长度泛化困难
  • 组合局限性:串并联组合无法生成所有DAG拓扑,即使考虑所有终端操作也无法捕获所有DP组合,揭示了结构表达的固有局限

行业启示

  • 当前神经网络架构在组合优化任务上的泛化瓶颈具有结构性根源,需从表示几何角度设计新的架构范式
  • 热带几何与代数几何工具可为AI模型的可解释性和泛化分析提供新的理论框架
  • 未来研究应关注如何突破DAG结构的组合限制,开发能更好捕捉动态规划内在几何特性的学习方法

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Programming 编程 LLM 大模型 Training 训练