ALCL: An Adaptive Log-Correntropy Loss for Robust Learning under Non-Gaussian Noise
Researchers have developed a new Adaptive Log-Correntropy Loss (ALCL) designed to improve the robustness of deep learning models when trained with non-Gaussian noise. Unlike traditional methods like mean squared error (MSE) that are sensitive to outliers, ALCL dynamically learns its robustness parameters during training. This adaptive approach, demonstrated on image datasets, consistently outperforms MSE and static correntropy losses, especially in high-noise conditions, by improving accuracy and reducing variance. AI
IMPACT Enhances deep learning model performance in noisy environments, potentially improving reliability in real-world applications.