神经形态系统中结构吸引子的形成
文章背景与核心概要
本文介绍了“不变结构学习”(Invariant Structural Learning,简称 ISL),这是一个针对神经形态及人工智能系统中概念形成的新型非优化学习框架。与依赖全局损失函数极小化和反向传播的传统方法不同,ISL 将学习过程构建为向超图空间中结构吸引子的收敛。
该研究在数学上对 ISL 模型进行了全面形式化,证明了其有限收敛性以及类结构吸引子的存在性和唯一性,并通过极少的训练数据在图像识别任务中验证了其计算可行性。此外,本文还探讨了其在树突树和神经回路中的前瞻性神经生物学实现,为理解大脑计算提供了崭新的理论视角。
执行摘要 (Executive Summary)
本文介绍了不变结构学习(Invariant Structural Learning,简称 ISL),这是一个用于神经形态和人工智能系统中概念形成的新型非优化框架。ISL 没有依赖全局损失函数的极小化和反向传播,而是将学习过程构建为向超图空间内结构吸引子的收敛。该研究提供了全面的数学形式化表达,在极少训练数据的图像识别任务上验证了其计算可行性,并探讨了其在树突树和神经回路中的潜在神经生物学实现。
This paper introduces Invariant Structural Learning (ISL), a novel non-optimization-based framework for concept formation in neuromorphic and artificial intelligence systems. Instead of relying on global loss function minimization and backpropagation, ISL frames learning as a convergence toward structural attractors within a hypergraph space. The study provides a comprehensive mathematical formalization, validates the computational feasibility on image recognition tasks with minimal training data, and explores prospective neurobiological implementations in dendritic trees and neural circuits.
元数据与文档信息 (Metadata & Document Information)
| 字段 | 详情 |
|---|---|
| arXiv ID | arXiv:2609.06826 [cs.AI] |
| 主要学科 | 人工智能 (cs.AI) |
| 次要学科 | 机器学习 (cs.LG);神经与进化计算 (cs.NE);神经元与认知 (q-bio.NC) |
| ACM 分类 | I.2.6 |
| 作者 | Yurii Parzhyn, Alexander Schwarzmann, Mykyta Lapin, Kostiantyn Bokhan |
| 提交日期 | 2026年9月6日 |
| 文档规格 | 131 页,4 个图表,5 个表格 |
Field Detail arXiv ID arXiv:2609.06826[cs.AI]Primary Subject Artificial Intelligence ( cs.AI)Secondary Subjects Machine Learning ( cs.LG); Neural and Evolutionary Computing (cs.NE); Neurons and Cognition (q-bio.NC)ACM Classification I.2.6 Authors Yurii Parzhyn, Alexander Schwarzmann, Mykyta Lapin, Kostiantyn Bokhan Submission Date September 6, 2026 Document Specs 131 pages, 4 figures, 5 tables
摘要 (Abstract)
本文探讨了不变结构学习(ISL)理论,该理论提出了一种用于概念形成的非优化方法。学习被解释为向超图空间中结构吸引子的收敛,而不是全局损失函数的最小化。
本文从三个主要维度呈现了 ISL 模型: 1. 数学形式化: 引入了结构归约过程的形式化工具,证明了其有限收敛性、类结构吸引子的存在性与唯一性,以及吸引子图谱的自组织性。 2. 计算验证: 展示了该方法在经典图像识别任务中的可行性,利用无需反向传播的学习机制,且仅需极小的训练数据集。 3. 神经生物学解释: 提出了关于在树突树中实现结构吸引子、将神经编码作为内部吸引子动力学投影,以及发展支持所提学习概念的神经架构的可检验假设。
这些神经生物学机制被构建为锚定在现代实验数据之上的可检验假设,涉及树突计算、突触可塑性以及神经回路组织。
This paper examines the theory of Invariant Structural Learning (ISL), which proposes a non-optimization approach to concept formation. Learning is interpreted as convergence to structural attractors in a hypergraph space, rather than as the minimization of a global loss function.
The paper presents the ISL model across three primary dimensions: 1. Mathematical Formalization: Introduces the formal apparatus of the structural reduction process, proving its finite convergence, the existence and uniqueness of class structural attractors, and the self-organization of attractor maps. 2. Computational Verification: Demonstrates the feasibility of the proposed approach on classical image recognition tasks, utilizing the learning mechanism without backpropagation and using extremely small training datasets. 3. Neurobiological Interpretation: Formulates testable hypotheses regarding the implementation of structural attractors in dendritic trees, neural coding as a projection of internal attractor dynamics, and the development of neural architectures supporting the proposed learning concept.
The neurobiological mechanisms are framed as testable hypotheses anchored in modern experimental data regarding dendritic computations, synaptic plasticity, and neural circuit organization.
核心贡献与主要章节 (Key Contributions & Core Sections)
1. 数学框架
- 结构归约: 为归约过程建立了一个严格的形式化工具。
- 收敛性与吸引子: 证明了有限收敛性,以及类结构吸引子的存在性和唯一性。
- 自组织: 展示了吸引子图谱如何在超图空间内实现自组织。
1. Mathematical Framework
- Structural Reduction: Establishes a rigorous formal apparatus for the reduction process.
- Convergence & Attractors: Formally proves finite convergence alongside the existence and uniqueness of class structural attractors.
- Self-Organization: Demonstrates how attractor maps self-organize within hypergraph spaces.
2. 计算实现
- 无反向传播: 绕过了传统的梯度下降和损失最小化范式。
- 低数据效率: 证明了在使用极小训练集的情况下,在经典图像识别基准上也能表现出有效性能。
2. Computational Implementation
- No Backpropagation: Bypasses conventional gradient descent and loss minimization paradigms.
- Low-Data Efficiency: Proves effective performance on classical image recognition benchmarks using extremely small training sets.
3. 神经生物学假设
- 树突计算: 探讨了结构吸引子如何在树突分支中物理表现出来。
- 神经编码: 将神经编码模式解释为内部吸引子动力学的投影。
- 实验相关性: 将理论发现与突触可塑性和结构电路拓扑的当代见解联系起来。
3. Neurobiological Hypotheses
- Dendritic Computation: Explores how structural attractors could physically manifest within dendritic arborizations.
- Neural Coding: Interprets neural coding patterns as projections of internal attractor dynamics.
- Experimental Relevance: Connects theoretical findings to contemporary insights in synaptic plasticity and structural circuit topology.
访问与资源 (Access & Resources)
- 全文 PDF: 通过 arXiv 查看 PDF
- 数字对象唯一标识符 (DOI): 10.48550/arXiv.2609.06826
- 替代版本: arXiv:2609.06826v1
Access & Resources
- Full-Text PDF: View PDF via arXiv
- Digital Object Identifier (DOI): 10.48550/arXiv.2609.06826
- Alternative Versions: arXiv:2609.06826v1