跳转至

主动推断生成模型的重整化：基础、推导与验证

文章背景与核心概要

主动推断（Active Inference）为感知、学习和行动提供了一个统一的框架，然而将离散的主动推断模型扩展到丰富的时空域仍然是一个重大挑战。重整化生成模型（RGMs）通过在多个空间和时间尺度上组合离散的生成模型来解决这一难题，有效地将低层状态和路径粗粒化为针对对象、事件和动作的高层原因。

然而，完全复现和适配这一框架一直受到紧凑的数学表述以及深度嵌入在专业软件环境中的实现细节的阻碍。本文提供了自包含的、面向推导的 RGM 阐述，以及一个开源且经过验证的实现。通过阐明底层理论——例如层次结构如何构建、信念和动作如何更新以及信息如何在各层之间流动——这项工作提高了透明度、可审计性和可复现性，为未来在机器学习基准测试上的定量评估奠定了坚实的基础。

Active inference offers a unified framework for perception, learning, and action, but scaling discrete active-inference models to rich spatial and temporal domains remains a significant challenge. Renormalising Generative Models (RGMs) address this hurdle by composing discrete generative models across multiple spatial and temporal scales, effectively coarse-graining lower-level states and paths into higher-level causes for objects, events, and actions.

However, fully reproducing and adapting this framework has been hindered by compact mathematical expositions and implementation details deeply embedded within specialized software environments. This paper provides a self-contained, derivation-oriented account of RGMs alongside an open, verified implementation. By clarifying the underlying theory—such as how hierarchies are structured, beliefs and actions are updated, and information flows across levels—this work improves transparency, auditability, and reproducibility, laying a solid foundation for future quantitative evaluations on machine learning benchmarks.

文档详情

Document Details

  • arXiv ID: arXiv:2608.09512 [cs.AI]

  • 发布日期: 2026年8月10日

    • Published Date: August 10, 2026

  • 篇幅: 25页，1幅图

    • Length: 25 pages, 1 figure

  • 会议收录: 被第7届主动推断国际研讨会（IWAI 2026）接受为长论文

    • Conference Acceptance: Accepted as a full paper at the 7th International Workshop on Active Inference (IWAI 2026)

作者

Authors

  • Karim Zaghw

    • Karim Zaghw

  • Andrew Pashea

    • Andrew Pashea

  • Marc Pritsch

    • Marc Pritsch

  • Wouter Nuijten

    • Wouter Nuijten

  • Karl Friston

    • Karl Friston

  • Lancelot Da Costa

    • Lancelot Da Costa

分类与元数据

Classification & Metadata

  • 主要学科: 人工智能 (cs.AI)

    • Primary Subject: Artificial Intelligence (cs.AI)

  • 次要学科: 计算机视觉与模式识别 (cs.CV)

    • Secondary Subjects: Computer Vision and Pattern Recognition (cs.CV)

  • ACM 分类: I.2.6; I.2.8; G.3

    • ACM Classes: I.2.6; I.2.8; G.3

  • DOI: 10.48550/arXiv.2608.09512

资源与补充材料

Resources & Supplementary Materials