Stochastic Gradient Descent in the Saddle-to-Saddle Regime of Deep Linear Networks

• URL: http://arxiv.org/abs/2604.06366v1
• Date: Tue, 07 Apr 2026 18:43:08 GMT
• Title: Stochastic Gradient Descent in the Saddle-to-Saddle Regime of Deep Linear Networks
• Abstract summary: gradient descent (SGD) noise on descent in deep linear networks is poorly understood.<n>We investigate the dynamics of SGD during training of DLNs in the saddle-to-saddle regime.<n>Our results establish that SGD noise encodes information about the progression of feature learning but does not fundamentally alter the saddle-to-saddle dynamics.
• Score: 74.46751089984072
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Deep linear networks (DLNs) are used as an analytically tractable model of the training dynamics of deep neural networks. While gradient descent in DLNs is known to exhibit saddle-to-saddle dynamics, the impact of stochastic gradient descent (SGD) noise on this regime remains poorly understood. We investigate the dynamics of SGD during training of DLNs in the saddle-to-saddle regime. We model the training dynamics as stochastic Langevin dynamics with anisotropic, state-dependent noise. Under the assumption of aligned and balanced weights, we derive an exact decomposition of the dynamics into a system of one-dimensional per-mode stochastic differential equations. This establishes that the maximal diffusion along a mode precedes the corresponding feature being completely learned. We also derive the stationary distribution of SGD for each mode: in the absence of label noise, its marginal distribution along specific features coincides with the stationary distribution of gradient flow, while in the presence of label noise it approximates a Boltzmann distribution. Finally, we confirm experimentally that the theoretical results hold qualitatively even without aligned or balanced weights. These results establish that SGD noise encodes information about the progression of feature learning but does not fundamentally alter the saddle-to-saddle dynamics.

Related papers

• Why SGD is not Brownian Motion: A New Perspective on Stochastic Dynamics [92.39053980710702]
  Gradient Descent (SGD) is commonly modeled as a Langevin process, assuming that minibatch noise acts as Brownian motion.<n>This approximation relies on a continuous-time limit and a sqrt(eta) noise scaling that does not match the discrete SGD update at finite learning rate.<n>We propose an alternative formulation of SGD as deterministic dynamics in a fluctuating loss landscape induced by minibatch sampling.
  arXiv  Detail & Related papers  (2026-05-21T15:50:40Z)
• On the Learning Dynamics of Two-layer Linear Networks with Label Noise SGD [93.70725920710208]
  We study the learning dynamics of gradient descent with label noise.<n>We highlight the critical role of label noise in driving the transition from the lazy to the rich regime.<n>We extend these insights to Sharpness-Aware Minimization (SAM), showing that the principles governing label noise SGD also apply to broader optimization algorithms.
  arXiv  Detail & Related papers  (2026-03-11T04:26:54Z)
• Momentum Does Not Reduce Stochastic Noise in Stochastic Gradient Descent [0.6906005491572401]
  In neural deep networks, gradient descent (SGD) with momentum is said to converge faster and have better generalizability than SGD without momentum.<n>In particular, adding momentum is thought to reduce this batch noise.<n>We analyzed the effect of search direction noise, which is noise defined as the error between the search direction and the steepest descent direction.
  arXiv  Detail & Related papers  (2024-02-04T02:48:28Z)
• Noise Balance and Stationary Distribution of Stochastic Gradient Descent [10.621129623557884]
  We show that the minibatch noise of SGD regularizes the solution towards a noise-balanced solution whenever the loss function contains a rescaling parameter symmetry.<n>Because the difference between a simple diffusion process and SGD dynamics is the most significant when symmetries are present, our theory implies that the loss function symmetries constitute an essential probe of how SGD works.<n>We then apply this result to derive the stationary distribution of gradient flow for a diagonal linear network with arbitrary depth and width.
  arXiv  Detail & Related papers  (2023-08-13T03:13:03Z)
• Machine learning in and out of equilibrium [58.88325379746631]
  Our study uses a Fokker-Planck approach, adapted from statistical physics, to explore these parallels. We focus in particular on the stationary state of the system in the long-time limit, which in conventional SGD is out of equilibrium. We propose a new variation of Langevin dynamics (SGLD) that harnesses without replacement minibatching.
  arXiv  Detail & Related papers  (2023-06-06T09:12:49Z)
• Doubly Stochastic Models: Learning with Unbiased Label Noises and Inference Stability [85.1044381834036]
  We investigate the implicit regularization effects of label noises under mini-batch sampling settings of gradient descent. We find such implicit regularizer would favor some convergence points that could stabilize model outputs against perturbation of parameters. Our work doesn't assume SGD as an Ornstein-Uhlenbeck like process and achieve a more general result with convergence of approximation proved.
  arXiv  Detail & Related papers  (2023-04-01T14:09:07Z)
• Stochastic Modified Flows, Mean-Field Limits and Dynamics of Stochastic Gradient Descent [1.2031796234206138]
  We propose new limiting dynamics for gradient descent in the small learning rate regime called modified flows. These SDEs are driven by a cylindrical Brownian motion and improve the so-called modified equations by having regular diffusion coefficients and by matching the multi-point statistics.
  arXiv  Detail & Related papers  (2023-02-14T15:33:59Z)
• The Limiting Dynamics of SGD: Modified Loss, Phase Space Oscillations, and Anomalous Diffusion [29.489737359897312]
  We study the limiting dynamics of deep neural networks trained with gradient descent (SGD) We show that the key ingredient driving these dynamics is not the original training loss, but rather the combination of a modified loss, which implicitly regularizes the velocity and probability currents, which cause oscillations in phase space.
  arXiv  Detail & Related papers  (2021-07-19T20:18:57Z)
• Asymmetric Heavy Tails and Implicit Bias in Gaussian Noise Injections [73.95786440318369]
  We focus on the so-called implicit effect' of GNIs, which is the effect of the injected noise on the dynamics of gradient descent (SGD) We show that this effect induces an asymmetric heavy-tailed noise on gradient updates. We then formally prove that GNIs induce an implicit bias', which varies depending on the heaviness of the tails and the level of asymmetry.
  arXiv  Detail & Related papers  (2021-02-13T21:28:09Z)
• Training Generative Adversarial Networks by Solving Ordinary Differential Equations [54.23691425062034]
  We study the continuous-time dynamics induced by GAN training. From this perspective, we hypothesise that instabilities in training GANs arise from the integration error. We experimentally verify that well-known ODE solvers (such as Runge-Kutta) can stabilise training.
  arXiv  Detail & Related papers  (2020-10-28T15:23:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.