AI Paper Atlas

Research map / Learning theory and algorithms / Kernels and regression theory

Towards Data-Algorithm Dependent Generalization: a Case Study on Overparameterized Linear Regression

Jing Xu, Jiaye Teng, Yang Yuan, Andrew C Yao

NeurIPS 2023 · Learning theory and algorithms › Kernels and regression theory · approach: Information-theoretic generalization bounds · 1 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
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/i, n=100, p=1000excess risk: 24.071 -> 2.399Improved
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/i^2, n=100, p=1000excess risk: 43.472 -> 0.214Improved
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/i^3, n=100, p=1000excess risk: 10.401 -> 0.077Improved
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/(i log(i+1)), n=100, p=1000excess risk: 89.922 -> 0.697Improved
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/(i log^2(i+1)), n=100, p=1000excess risk: 82.413 -> 0.298Improved
use optimal early-stopping iterate instead of min-norm interpolatorGD on overparameterized linear regression · eigenvalues 1/(i log^3(i+1)), n=100, p=1000excess risk: 38.145 -> 0.187Improved
use best GD iterate instead of final SGD iterateCNN · Corrupted MNIST 20% label noisetest error (%): 19.9 -> 2.88Improved
faster eigenvalue decay of feature covariance (1/i to 1/i^3)early-stopped GD · synthetic linear regressionoptimal excess risk: 2.399 -> 0.077Improved
faster decay of eigenvaluesmin-norm interpolator · synthetic linear regressionexcess risk: 24.071 -> 43.472 -> 10.401Mixed
log-polynomial decay with larger power beta (1 to 3)early-stopped GD · synthetic linear regressionoptimal excess risk: 0.697 -> 0.187Improved

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