.png){kind=icon}
.png){kind=icon}
Return
Sample-customized implicit semantic data augmentation for neural networks regularization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.130288.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Data augmentation techniques have been extensively proven to have the remarkable capability of improving the generalization of deep neural networks via appropriate training data transformations. However, current research mainly focus on automatically learning the augmentation policy in a simply designed search space (e.g., typical pixel space geometric transformations or their combinations) for the whole training set, while overlooking the individual sample variations. As a result, coarse dataset-level augmentation policies with minimal transform operations are learned and applied to the training data, potentially introducing noisy samples and adversely impacting network training. In this paper, we propose a novel sample-level data augmentation method, named SCAug, by exploiting the sample-customized implicit semantic transformations as well as the adversarial learning scheme to enable more accurate and diversified data augmentation. Specifically, SCAug introduces a policy network to implicitly explore the proper augmentation policy in deep feature space with the consideration of individual sample features, and the augmentation policy learning procedure acts as one side of the adversarial learning aims to enrich the data diversity semantically and make the machine learning task more challenging. While the adversary, i.e., the task network, is trained to well tolerate these customized augmented samples, thereby becoming more robust and generalizable. Compared with existing methods, SCAug achieves superior performance on typical benchmarks including ImageNet, CIFAR and SVHN.
Keywords:
Data augmentation
Machine learning
Data transformation
Semantic transformation
Adversarial learning
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
View more organizations