AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Gradient Descent with Large Step Sizes: Chaos and Fractal Convergence Region

Shuang Liang, Guido Montufar

ICLR 2026 · Optimization › Implicit bias and generalization · approach: Implicit regularization of SGD

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add l2 regularization to scalar factorization loss (ablation)Gradient descent with near-critical step size · Scalar factorization convergence regionboundary regularity: smooth boundary -> fractal boundaryHurt
Use near-critical step sizes instead of small step sizesGradient descent · Depth-3 ReLU network on 2-class CIFAR-10 subset (MSE)predictability of final norm and sharpnessHurt
Use step sizes in the edge-of-stability regimeGradient descent · Depth-3 ReLU network on CIFAR-10 subsetfinal sharpness: sharpness near 2/etaNo clear effect
Add Polyak momentumGradient descent · Depth-3 ReLU network on CIFAR-10 subsetfinal sharpnessMixed
Add weight decay (ablation)Gradient descent · Depth-3 ReLU network on CIFAR-10 subsetfractal convergence structureNo clear effect
Use cross-entropy loss instead of MSE (ablation)Gradient descent · Depth-3 ReLU network on CIFAR-10 subsettwo-regime behaviorNo clear effect
Initialize in orthogonal subspace W (incl. identity init)Gradient descent matrix factorization · General matrix factorizationchaos and fractal boundaryNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy