Enhancing Audio Augmentation Methods with Consistency Learning

Data augmentation is an inexpensive way to increase training data diversityand is commonly achieved via transformations of existing data. For tasks suchas classification, there is a good case for learning representations of thedata that are invariant to such transformations, yet this is not explicitlyenforced by classification losses such as the cross-entropy loss. This paperinvestigates the use of training objectives that explicitly impose thisconsistency constraint and how it can impact downstream audio classificationtasks. In the context of deep convolutional neural networks in the supervisedsetting, we show empirically that certain measures of consistency are notimplicitly captured by the cross-entropy loss and that incorporating suchmeasures into the loss function can improve the performance of audioclassification systems. Put another way, we demonstrate how existingaugmentation methods can further improve learning by enforcing consistency