AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

How SGD Selects the Global Minima in Over-parameterized Learning: A Dynamical Stability Perspective

Lei Wu, Chao Ma, Weinan E

NeurIPS 2018 · Optimization › Implicit bias and generalization · approach: SGD noise and flat minima · 69 citations (OpenAlex)

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
decrease batch size from GD (1000) to small SGD batch (ablation)SGD on CIFAR10 VGG · CIFAR10 non-uniformity, eta=0.01non-uniformity: about 350 -> about 100Improved
increase learning rate (ablation)GD · FashionMNIST and CIFAR10sharpness of found minimum: 53.5 -> 1.9 (FashionMNIST, eta 0.01 -> 1)Improved
use SGD with small batch instead of GD (ablation)SGD · FashionMNIST and CIFAR10sharpnessImproved
switch from GD eta=0.1 to SGD eta=0.1 B=4corrupted FashionMNIST FNN · Corrupted FashionMNISTtest accuracy: 71.44 -> 72.42Improved
switch from GD eta=0.1 to SGD eta=0.1 B=4FashionMNIST FNN · FashionMNISTtest accuracy: 80.04 -> 80.40Improved
use batch size 100 instead of 4 (SGD eta=0.1) (ablation)SGD · Corrupted FashionMNISTescape from minimum: 70.90 vs 72.42 test accHurt
run GD/SGD from an L-BFGS minimumL-BFGS solution · Corrupted FashionMNISTtest accuracy: 69.5 -> 73.25Improved
flatter minima vs sharperminima selected · Clean FashionMNISTtest accuracy correlation with sharpnessMixed

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