AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Cheap Bootstrap for Fast Uncertainty Quantification of Stochastic Gradient Descent

Henry Lam, Zitong Wang

JMLR 2026 · Optimization › Implicit bias and generalization · approach: SGD noise and flat minima

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
increase number of resamples B from 3 to 10 (ablation)COfB ASGD · Linear regression, Toeplitz, n=1e5, d=200average interval length (x1e-2): 14.44 -> 8.44Improved
increase number of resamples B from 3 to 10 (ablation)COnB · Linear regression, Toeplitz, n=1e5, d=200average interval length (x1e-2): 21.04 -> 15.61Improved
increase B from 3 to 10 (ablation)COfB/COnB · interval width vs normalt-interval relative width: 49.6% wider -> 10.9% widerImproved
increase B from 3 to 10 (ablation)COfB ASGD · Linear regression, Toeplitz, n=1e5, d=200coverage probability (%): 94.41 -> 94.01No clear effect
increase B from 3 to 10 in sparse regression (ablation)COfB ASGD · Sparse linear regression p=3, in T*average interval length (x1e-2): 2.75 -> 0.86Improved
use averaging (ASGD) instead of last-iterate SGD (ablation)COfB B=3 · Linear regression, Toeplitz, n=1e5, d=5average interval length (x1e-2): 6.93 -> 3.06Improved
use cheap bootstrap B=3 instead of delta method (ablation)SGD CI · Linear regression, Toeplitz, n=1e5, d=200coverage probability (%): 37.16 -> 94.41Improved
increase dimension from 20 to 200delta method · Linear regression, Toeplitz, n=1e5coverage probability (%): 93.23 -> 37.16Hurt
use cheap bootstrap B=3 instead of HiGrad (ablation)SGD CI · Logistic regression, Toeplitz, n=1e5, d=200coverage probability (%): 57.02 -> 93.79Improved
reduce online bootstrap B from 100 to 10 (ablation)online bootstrap · Linear regression, Toeplitz, n=1e5, d=200coverage probability (%): 96.58 -> 93.33No clear effect
use COnB B=3 instead of online bootstrap B=100 (ablation)SGD CI · Linear regression runtime d=10runtime (seconds): 45.90 -> 1.75Improved

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