ICLR

Weak Correlations as the Underlying Principle for Linearization of Gradient-Based Learning Systems

弱相关性作为基于梯度的学习系统线性化的底层原理

Ori Shem-Ur, Khen Cohen, Aviv Orly, Yaron Oz

AI总结 该研究以宽神经网络等深度学习模型为对象，揭示基于梯度下降的学习系统的线性化源于假设函数导数间的弱相关性，推导了随机梯度下降训练的线性偏离界，并提出了表征随机张量渐近行为的新方法。

Comments 41 pages; 10 pages main tex; 0 figures. Aviv Orly added new results requested by the ICLR reviewers after the discussion phase had ended

Journal ref International Conference on Learning Representations (ICLR), 2026