AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

When Sample Selection Bias Precipitates Model Collapse

Xinbao Qiao, Xianglong Du, Wei Liu, Jingqi Zhang, Peihua Mai, Meng Zhang, Yan Pang

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 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
select synthetic data against a skewed local reference (Airplane class only) with CenterMatchrecursive DDPM training on CIFAR-10 · CIFAR-10, Airplane-only local referenceclass diversityHurt
use local-reference selection baselines instead of random selection under non-IID referenceAccumulate-Subsample recursive DDPM · CIFAR-10, ExDir(1, 0.1) referenceFID: 106 -> 116 (CenterMatch), 106 -> 115 (CovMatch)Hurt
use K-means selection instead of randomrecursive DDPM · CIFAR-10, ExDir(1, 0.1) referenceFID: 106 -> 102No clear effect
use local-reference selection baselines in the IID partitionrecursive DDPM · IID partition referenceFIDImproved
use collaborative geodesic interpolation proxy (Scheme I) instead of random selectionrecursive DDPM · CIFAR-10, ExDir(1, 0.1) referenceFID: 106 -> 71Improved
use Wasserstein barycenter proxy (Scheme II) instead of random selectionrecursive DDPM · STL-10, ExDir(1, 0.1) referenceFID: 95 -> 69Improved
use Scheme I instead of Scheme IIcollaborative proxy selection · CIFAR-10, ExDir(1, 0.1) referenceFID: 85 -> 71Improved
use Scheme II (reusable barycenter) instead of Scheme I as number of clients growscollaborative proxy selection · selection wall-clock time, K=100time (s): 49.2 -> 42.1Improved
add top-alpha selection toward a single local target to accumulate paradigmGaussian self-consuming loop · multivariate Gaussian simulationTr(Sigma_t)/Tr(Sigma_0)Hurt
selection baselines on highly structured face data (CelebA) with a biased referenceCenterMatch/CovMatch/K-means · CelebA, ExDir(1, 0.1) referenceFID: 96 -> 87 (K-means, CenterMatch)Improved

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