Modeling Uncertainty by Integrating Label Smoothing and Data Perturbation: An Application to Pneumonia Detection

The article presents an integrated learning methodology that combines label smoothing regularization, noise augmentation, and a warm-up training strategy for deep neural networks, aiming to enhance model generalization and stability. Label smoothing regularization is employed to introduce soft decision making in model’s predictions by incorporating a small amount of softness into the target labels during training. This technique effectively prevents the model from becoming overly reliant on hard binary labels, thus improving its ability to generalize to unseen data. To simulate real-world variability, perturbation in imaging data is introduced during training. Furthermore, a warm-up training strategy is implemented to gradually adjust the model parameters at the beginning of training. This strategy helps in stabilizing the convergence process, allowing the model to explore the parameter space more effectively and avoid premature convergence to suboptimal solutions. Experimental evaluations are conducted on a dataset of chest X-ray images to validate the efficacy of the proposed framework. Results demonstrate that the integrated approach yields superior accuracy of 97.02% as compared to state-of-the-art methods.