Noisy labels
Noisy labels: Incorrect or inconsistent labels in data that can distort learning and evaluation. Noise can be random, class-dependent or systematic with respect to features. Clean validation samples, robust losses and human review help distinguish it from genuinely ambiguous cases.
In italiano: Noisy LabelsLegge III · Come una rete impara dai datiNoise can be random, class-dependent or systematic with respect to features. Clean validation samples, robust losses and human review help distinguish it from genuinely ambiguous cases.
Examples
- Conflicting crowdsourced labels
- Classes assigned by heuristic rules
- Errors in a validation set