ConceptLabel Noise Tolerance
Label Noise Tolerance
The ability of a neural network to tolerate incorrect labels in training data. Deep networks preferentially fit clean sample patterns during early learning, so when the label error rate is below a certain threshold, the degradation in generalization performance remains limited.
Timeline (last 90 days)
Oct 15
When the label error rate is below about 10%-20%, deep networks can still learn effective representations, and the degradation in generalization performance is limited.
Unverified50%