新研究将SGD动力学建模为渗透过程

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-03 04:00

研究人员将随机梯度下降（SGD）的动力学建模为渗透过程，揭示了架构对称性如何导致子网络以离散块的形式合并。这些转变表现为宏观序参量中的方差峰值，类似于物理相变。研究表明，这种捕获机制及其相关的尺度级联也适用于重尾噪声模型下的Adam和AdamW优化器。 AI

影响为理解深度学习中的优化动力学提供了一个新的理论框架。

排序理由该集群包含一篇发表在arXiv上的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 23 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该集群包含一篇发表在arXiv上的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, other; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Breaking (< 6h); Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

arXiv cs.AI TIER_1 English(EN) · Sai Niranjan Ramachandran, Suvrit Sra · 2026-09-03 04:00

优化中的渗流动力学：方差级联与离散尺度不变性

arXiv:2609.02373v1 Announce Type: cross Abstract: We study the dynamics of Stochastic Gradient Descent (SGD), which is known to steer deep neural networks toward invariant sets that correspond to simpler subnetworks. How this steering unfolds over time remains poorly understood. …