This thesis investigates the effectiveness of adaptive label smoothing in enhancing model generalization and robustness within image classification tasks that utilize aggressive augmentations from automated strategies. Traditional augmentation techniques assign fixed labels regardless of transformation severity, often leading to overconfident models. Inspired by SoftAugmentation (SA)---a method that significantly improved model performance on occluded images by addressing this limitation through a visibility-aware approach based on Human Visual System (HVS) studies and dynamically adjusts label confidence according to the extent of transformation-induced information loss, particularly in scenarios involving aggressive image cropping.---this research extends adaptive label smoothing to a broader spectrum of augmentations using TrivialAugment (TA), an automated augmentation strategy. The study introduces four adaptive label smoothing approaches that dynamically adjust label confidence based on augmentation intensity: reference model performance, HVS accuracy curves, image comparison metrics, and polynomial smoothing functions. These methods are evaluated on diverse datasets including CIFAR-10, CIFAR-100, and Tiny-ImageNet, to assess their effectiveness.
Experimental results indicate that adaptive label smoothing significantly enhances model performance when applied to individual aggressive augmentations. However, its effectiveness diminishes when augmentations are applied uniformly through automated strategies like TA, where the combined effects of multiple augmentations limit its impact. This limited improvement could be due to factors such as the effectiveness of augmentations varies significantly across different classes and, there is a lack of comprehensive HVS studies for many augmentation types. These findings highlight the need for further research into more sophisticated adaptive label smoothing approaches as the label adjustment methods are highly augmentation-dependent, to improve model performance.