Intermix: An Interference-Based Data Augmentation and Regularization Technique for Automatic Deep Sound Classification

In this paper, we present InterMix, an interference-based regularization and data augmentation strategy for automatic sound classification. InterMix creates virtual training examples by creating an interference-based mixed representation for a sampled phase difference and mixup ratio. InterMix can be used to train sound classification models with the ability to generate a vast amount of training samples. These are significantly varied compared to that of other mixup strategies due to the introduction of phase difference, a continuous variable. While building on other mixup strategies which use linear interpolation, we perform mixup based on the formula of interference. We demonstrate the utility of InterMix in comparison to standard learning techniques and previously applied mixing strategies through a quantitative analysis. We also demonstrate that InterMix is more robust towards adversarial attacks compared to standard learning and other mixup strategies.