Data Driven Augmentation for Deep Learning Applications

Abstract: Deep learning models are achieving remarkable performance on numerous tasks across various fields and applications. However, current deep learning models often suffer from overfitting and are therefore heavily reliant on regularization techniques such as data augmentation. While data augmentation is widely employed in deep learning applications, recent research has shown that it introduces a class bias during training of image classification models, affecting the performance of classes differently.     This study aims to investigate the class bias introduced by data augmentation in deep learning models. Through an empirical study, we examine the effects of popular augmentation techniques, namely random resized crop, Gaussian blur and color jitter, on both the overall model performance and the performance of individual classes. Additionally, potential benefits of class specific data augmentation are studied. The findings confirm that data augmentation indeed introduces a class bias and we observe that the introduction of data augmentation leads to extensive performance reduction for some classes. Surprisingly, we find that augmenting other classes might have a greater impact on the performance of a specific class than augmentation applied directly to that class. The results from this study point to the need for careful consideration when employing data augmentation techniques in order to ensure the development of high quality models.