What question did this study set out to answer?

The study aims to compare pooled cluster posteriors against local k-nearest-neighbour estimators in detecting label noise.

July 29, 2026Open Access

Model-independent label-noise detection by feature geometry: pooled vs. local estimators

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial compares pooled and local estimators for detecting noisy labels, suggesting pooled methods excel under specific conditions.

The study aims to compare pooled cluster posteriors against local k-nearest-neighbour estimators in detecting label noise.
Utilized Confident Learning to detect label errors by training classifiers on noisy labels.
Evaluated local k-nearest-neighbour and pooled bagged cluster posteriors using feature-space geometry.
Conducted tests on real human noise data (CIFAR-10N) with varying levels of feature-dependent noise (>=40%).
SimiFeat showed strengths overall, but pooling outperformed it under high feature-dependent noise (F1 score 0.930 vs. 0.915).
Pooling the geometry into bagged cluster posteriors provides better performance than local k-NN in corrupted local neighborhoods.
A per-class voting rule for geometric posteriors was more effective than Confident-Learning pruning.

Daniel Hough (2026) studied this question.

View Full Paper

Ask AI

Bookmark

Randomized trial compares pooled and local estimators for detecting noisy labels, suggesting pooled methods excel under specific conditions.