Georg Siedel

Abstract

Soft augmentation regularizes the supervised learning pro- cess of image classifiers by reducing label confidence of a training sample based on the magnitude of a Random Crop augmentation applied to it. This paper extends this adaptive label smoothing framework to other types of aggressive augmentations beyond Random Crop. Specifically, we demonstrate the effectiveness of the method for random erasing and noise injection data augmentation. Adaptive label smoothing permits stronger regularization via higher-intensity Random Erasing. However, its benefits vanish when applied with a diverse range of image transfor- mations as in the state-of-the-art TrivialAugment method, and excessive label smoothing harms robustness to common corruptions. Our findings suggest that adaptive label smoothing should only be applied when the training data distribution is dominated by a limited, homogeneous set of image transformation types.

Abstract

Abstract Robustness is a vital characteristic of machine learning classifiers that is required to ensure safety and reliability. In the extensively studied domain of adversarial robustness in image classification, it is typically defined as a model’s stability in response to input perturbations within a p -norm distance. However, within the research area focused on robustness against random corruptions, p -norm perturbations are rarely investigated, with an emphasis instead on real-world corruptions. This paper examines the application of random p -norm corruptions to enhance the training and testing datasets of image classifiers. We evaluate model robustness against subtle random p -norm corruptions and propose an associated robustness metric. Through empirical analysis, we investigate whether robustness is transferable across different p -norms and identify which p -norm corruptions are most effective for training and evaluation. Furthermore, we introduce efficient training data augmentation techniques that combine various p -norm corruptions and assess their effectiveness against several state-of-the-art noise injection methods. Our results indicate significant improvements in corruption robustness when utilizing these training approaches, even when combined with other prevalent data augmentation strategies.