Nature Machine Intelligence:分形盆地几何解释深度学习可预测性与可复现性的局限
摘要
Ly 与 Gong 在 Nature Machine Intelligence 发表研究,提出分形盆地几何是深度学习结果难以预测和复现的机制:通向不同结果的起点在任意精细尺度上交织,因此提高精度带来的改善有限。论文于 2026 年 11 月 5 日在线发表,DOI 为 10.1038/s42256-026-01336-3。
摘要由模型根据原文撰写,以原文为准。
正文 · 原文
Nature Machine Intelligence, Published online: 05 November 2026; doi:10.1038/s42256-026-01336-3
Deep learning outcomes are hard to predict and reproduce. Ly and Gong propose fractal basin geometry as a mechanism: starting points leading to different outcomes interweave at arbitrarily fine scales, so greater precision yields little improvement.
注意转载授权 · 阅读原文
来源:Nature Machine Intelligence · nature.com
引用原文
Andrew Ly. Fractal basin geometry and the limits to predictability and reproducibility in deep learning[EB/OL]. (2026-11-05)[访问日期]. https://www.nature.com/articles/s42256-026-01336-3.