Hierarchical Mixup: Improving Representations By Binary Masks And Interpolating Hidden States

Mixed Sample Data Augmentation (MSDA) methods are powerful Data Augmentation that interpolates between two or more examples and their corresponding target labels in the input or feature space. Recently, the Mixup method overlays or cuts and pastes two or more objects into one image, but this requires attention to the selection region. To address these issues, we propose the Hierarchical Mixup method, a simple MSDA method that enables a neural network to make predictions about the interpolation of hidden representations. It utilizes the mixed, augmented features as additional training samples to obtain a neural network that performs better on multiple hidden representations. Using this method allows the neural network to learn classes with less data. In addition, Hierarchical Mixup uses random binary masks obtained by applying thresholds to low-frequency images sampled from the Fourier space. It uses the generated binary masks as weights for MSDA. These random masks can present various shapes and be generated for two-dimensional data. Hierarchical Mixup is a lightweight module that can be applied to existing classification models. In this study, we used PreAct ResNet-18 as the base model, and our method achieved accuracies of 97.35% and 81.14% on the CIFAR-10/100 datasets, respectively. The experiments demonstrate that Hierarchical Mixup enhances the generalization ability of Deep Neural Networks (DNNs) and shows significant improvements in different datasets.