Vicinal Data Augmentation for Classification Model via Feature Weaken

Songhao Jiang, Yan Chu, Tianxing Ma, Xiaochen Miao, Zhengkui Wang, Tianning Zang

Abstract: Deep classification is the foundation task of knowledge science and has always been a focus of research. Recently, data augmentation technology has shown good advantages in enhancing model performance in classification tasks. We propose a novel training and data augmentation method, Feature Weaken, which can achieve limitations on model weight and generate a vicinal distribution of data to improve model performance and generalization ability. This enables the model to be suitable for weakened environments, thereby enhancing its decision-making ability in natural feature environments (without the weakened features). Our extensive evaluations over both image and text datasets show that Feature Weaken outperforms the baselines significantly w.r.t. the model performance. It is also proved that Feature Weaken can work together with other data augmentation techniques for further improvement.