Dissecting Mixed-Sample Data Augmentation Models via Neural-concept Association

Soyoun Won, Sung-Ho Bae, Seong Tae Kim

Abstract: Data augmentation techniques are widely employed in the training of deep neural networks (DNNs), and recent research verifies their effectiveness across diverse tasks. However, their impact on the model’s ability to capture semantic concepts has not been widely investigated. In this paper, we analyze models trained with various mixed-sample data augmentation strategies in terms of neural-concept association. Experimental results suggest that mixed sample data augmentation strategies make the model less reactive to semantic concepts.