返回全部动态
预条件指数转负：学习率耦合与跨环境泛化
原标题：When the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization
AI 摘要
该研究通过四环境配对分类任务，系统扫描了21个预条件指数和5个学习率，共420次训练。发现最大化跨环境准确率的预条件指数随学习率增大而近似线性下降，大学习率下最优指数进入负值区间。源域验证选择的指数系统性地高于跨环境最优值，揭示了分布偏移下的模型选择冲突。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
When the Preconditioning Exponent Turns Negative: Learning-Rate Coupling and Cross-Environment Generalization Abstract Adaptive optimizers are commonly parameterized by a fixed power of the second-moment estimate. Existing partially adaptive methods study exponents between momentum-like updates and the standard Adam square root, while the interaction between this exponent and the global learning rate is less understood. We perform a controlled cross-environment study using a paired four-environm
发布时间：2026-09-28 12:00
抓取时间：2026-09-28 12:00
来源机构：arXiv