Adaptive Generalized Cross-Entropy Loss for Sound Event Classification with Noisy Labels

Considering the high cost of manually annotated large-scale datasets for superior sound event classifier performance, the data collection process has shifted to using the Internet, which facilitates easier user-contributed audio and metadata collection. However, label noise is inevitable. To address the problems caused by label noise, several types of noise-robust loss functions have been proposed recently as alternatives to the commonly categorical cross-entropy (CCE) loss, one of which is the generalized cross-entropy (GCE) loss, which demonstrates state-of-the-art performance. However, GCE cannot realize sufficient noise robustness and satisfactory accuracy simultaneously. Thus, we propose adaptive GCE loss, which automatically adapts to noisy labels in every batch to achieve adequate noise robustness and sufficient accuracy. We conducted experiments and found that the classification accuracy of the proposed loss demonstrated 4.7% and 1.2% absolute improvement over the CCE and GCE baselines, respectively. We also demonstrate that clean data consumption in the proposed loss is dramatically reduced by more than 75% compared with CCE.