AI Paper Atlas

Research map / Label-efficient and robust learning / Noisy labels

Linear Distance Metric Learning with Noisy Labels

Meysam Alishahi, Anna Little, Jeff M. Phillips

JMLR 2024 · Label-efficient and robust learning › Noisy labels · approach: Contrastive learning on noisy data · 0 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 covariance normalization on the combustion data instead of scaling-only normalization (ablation)Logistic metric learning · Combustion equations-of-staterelative spectral error: 0.748 -> 0.041Improved
add covariance normalization to the input data (ablation)Logistic metric learning · Combustion equations-of-statemean test accuracy: 99.51% -> 99.73%Improved
feed raw unscaled data to the solver (ablation)Logistic metric learning · Combustion equations-of-stateconvergenceHurt
use noise-matched (Logistic) labels instead of direct random label flipping (ablation)Logistic loss model · Synthetictest accuracy w/o noise: 98.83% -> 94.51%Improved
use noise model close to Logistic rather than HS noise (ablation)Logistic loss model · Synthetictest accuracy w/ noise: 89.85% -> 85.28%Improved
increase mislabeling from 10% to 45% while holding training size (ablation)Logistic loss model · Synthetic, Logistic noiseaccuracy vs ground truth: around 95% at 40% noise; collapses at 45%Hurt
increase training set to 2e5 samples under 45% noise (ablation)Logistic loss model · Synthetic, 45% Logistic noisetest accuracy vs ground truth: 65% -> 97%Improved
increase number of training points from 100 to about 1000 (ablation)Logistic loss model · Synthetictrain/test accuracy gapImproved
use the proposed Logistic loss model instead of DML-eigMetric learning · Synthetic, Logistic noisetest accuracy vs ground truth: 90% -> 99%Improved
use the proposed Logistic loss model instead of DML-eigMetric learning · Synthetic training timetraining time: 1200 seconds -> less than a secondImproved
replace Logistic loss with Laplace or HS loss under Logistic noise (ablation)Metric learning loss · Synthetic eigenvalue recoveryeigenvalue recoveryNo 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