arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2606.04476cs.LGmath.OCmath.STstat.MLstat.TH
收藏

当两层都学习:通过ReLU网络表示线性模型的训练动力学

When Both Layers Learn: Training Dynamics of Representing Linear Models via ReLU Networks

Department of Electrical and Computer Engineering(电气工程系);Department of Computer Science(计算机科学系)

Berk Tinaz, Changzhi Xie, Mahdi Soltanolkotabi

AI总结:

本文研究单隐层ReLU网络联合训练两层以拟合线性目标函数的梯度下降动力学,通过三阶段分析证明从随机初始化出发能以线性速率收敛到全局最小化器并达到最优样本复杂度。

摘要:

在本文中,我们研究了联合训练单隐层ReLU网络的两层以拟合线性目标函数的梯度下降动力学。具体来说,我们考虑一个可实现设置,其中输入从高斯分布中独立同分布采样,标签遵循一个植入的线性模型。这种风格化的框架捕捉了逆问题和某些自编码器模型中端到端训练的关键特征。尽管其表面简单,但动力学仍然难以理解,部分原因是损失景观包含多个非严格鞍点,这使得不清楚为什么从随机初始化开始的梯度下降能够可靠地逃离坏的驻点区域。我们提供了优化景观的详细刻画,并证明从适度小的随机初始化开始——同时训练两层——梯度下降以线性速率收敛到全局最小化器,并具有阶次最优的样本复杂度。我们的分析通过三个阶段追踪轨迹:对齐阶段,其中隐藏权重逐渐与植入方向对齐,而输出权重保持正确的符号模式;增长阶段,其中两层的范数增加同时保持对齐;以及局部细化阶段,其中对齐的神经元快速收敛到植入方向,产生快速的局部收敛。为了严格证明梯度下降避免非严格鞍点,我们为端到端动力学开发了轨迹级控制论证。此外,我们建立了沿整个轨迹成立的新颖的均匀集中结果,这对于获得阶次最优的样本复杂度至关重要。我们通过一系列配置的大量实验验证了我们的理论。

英文摘要:

In this paper, we study the gradient descent dynamics for jointly training both layers of a one-hidden-layer ReLU network to fit a linear target function. Concretely, we consider a realizable setting where inputs are drawn i.i.d. from a Gaussian distribution and labels follow a planted linear model. This stylized framework captures salient features of end-to-end training in inverse problems and certain auto-encoder models. Despite its apparent simplicity, the dynamics remain poorly understood, in part because the loss landscape contains multiple non-strict saddle points, making it unclear why gradient descent from random initialization reliably escapes bad stationary regions. We provide a detailed characterization of the optimization landscape and prove that gradient descent from a moderately small random initialization-simultaneously training both layers-converges to a global minimizer at a linear rate with order-wise optimal sample complexity. Our analysis tracks the trajectory through three phases: an alignment phase in which hidden weights progressively align with the planted direction while the output weights maintain the correct sign pattern; a growth phase in which the norms of both layers increase while preserving alignment; and a local refinement phase in which the aligned neurons rapidly converge to the planted direction, yielding fast local convergence. To rigorously show that GD avoids non-strict saddles, we develop trajectory-level control arguments for the end-to-end dynamics. In addition, we establish novel uniform concentration results that hold along the entire trajectory, and are essential for obtaining order-wise optimal sample complexity. We corroborate our theory with extensive experiments across a range of configurations.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑