Browse Teach Curate

T-Revision-Alpha and T-Revision-Softmax for Label-Noise-Robust Classification

Two enhancements to the T-Revision method for learning classifiers robust to label noise: T-Revision-Alpha and T-Revision-Softmax improve the stability and robustness of noise-transition-matrix estimation under the anchor-point assumption, evaluated against forward-correction and importance-reweighting baselines on FashionMNIST and CIFAR-10.

Demerzel Text

Best Transition Matrix Esitimation or Best Label Noise Robustness Classifier? Two Possible Methods

This report studies learning image classifiers that are robust to label noise (incorrect training labels). It frames the problem through the noise transition matrix relating clean to observed labels.…