Adaptive Data Augmentation: Elliptical Masking and Dynamic Regularization with Feedback

In recent years, machine learning (ML) models have achieved robust performance and effective generalization in real-world applications. While ML models often perform well on training data, they can struggle with unseen inputs due to overfitting. Data augmentation is a key strategy to increase data diversity and improve model adaptability. This paper introduces the Adaptive Data Augmentation (ADA) method with the newly introduced random Elliptical Cutout, a novel approach that dynamically adjusts augmentation parameters in an adaptive manner based on model performance. Unlike traditional static augmentation methods, ADA adaptively modifies key training parameters, such as batch size, dropout rate, and weight decay, in response to validation accuracy. When validation accuracy stagnates over multiple epochs, these parameters are automatically updated to enhance generalization and reduce overfitting. We validate our approach on diverse datasets, including CIFAR-10, X-RAY, and a Brain MRI, demonstrating significant improvements in model accuracy and robustness compared to well known MixUp method. The ADA method achieves Area Under the Receiver Operating Characteristic Curve (AUC-ROC) ranging from 98.19% to 99.85% across all datasets and models, effectively reducing overfitting while enhancing generalization and classification robustness.