Hard Cases, Bad Labels: Testing Error Exposure and Error Location in Uncertainty Sampling Under Bounded Label Noise
Not provided in the abstract
Abstract
This study investigates the performance of uncertainty sampling in active learning under various noise conditions and its impact on labeling efficiency.
Reality Card
Uncertainty sampling improved normalized balanced-accuracy area under the learning curve by 1.09 to 1.77 percentage points on all datasets under clean labels.
Uncertainty sampling showed a balanced accuracy improvement across datasets, with specific metrics indicating varying robustness depending on dataset and noise structure.
The robustness of uncertainty sampling is dependent on dataset, budget, noise structure, and evaluation metric, which may limit generalizability.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.