Data Augmentation Revisited: Rethinking the Distribution Gap between Clean and Augmented Data

By Zhuoxun He, Lingxi Xie, Xin Chen, Ya Zhang, Yanfeng Wang, Qi Tian. 2019

70 citations recorded by the source corpus.

Data augmentation has been widely applied as an effective methodology to improve generalization in particular when training deep neural networks. Recently, researchers proposed a few intensive data augmentation techniques, which indeed improved accuracy, yet we notice that these methods augment data have also caused a considerable gap between clean and augmented data. In this paper, we revisit this problem from an analytical perspective, for which we estimate the upper-bound of expected risk using two terms, namely, empirical risk and generalization error, respectively. We develop an understanding of data augmentation as regularization, which highlights the major features. As a result, data augmentation significantly reduces the generalization error, but meanwhile leads to a slightly higher empirical risk. On the assumption that data augmentation helps models converge to a better region, the model can benefit from a lower empirical risk achieved by a simple method, i.e., using less-augmented data to refine the model trained on fully-augmented data. Our approach achieves consistent accuracy gain on a few standard image classification benchmarks, and the gain transfers to object detection.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Albumentations: fast and flexible image augmentations (2018, background): “Reducing the amount of augmentations in the last few epochs of training was proposed as a way to reduce a distribution gap between clean and augmented data and to improve model’s performance [38].”
Circumventing Outliers of AutoAugment with Knowledge Distillation (2020, background): “We noticed prior work [15] argued that data augmentation may introduce uncertainty to the network training process because the training data distribution is changed, and proposed to switch off data augmentation at the end of the training stage to alleviate the empirical risk of optimization.”
Tradeoffs in Data Augmentation: An Empirical Study (2021, background): “Fine-tuning on clean data can improve ﬁnal accuracy (He et al., 2019).”
FMix: Enhancing Mixed Sample Data Augmentation (2020, methodology): “Although He et al. (2019) noted the importance of characterising the effect of data augmentation from an information perspective, they did not explore any measures that do so.”
A Fourier-based Framework for Domain Generalization (2021, methodology): “Similar as in [3, 13], we expand R̂aug with Taylor expansion:”