Abstract: For image classification tasks, machine learning models perform best when balancing meaningful feature representation with robustness to non-meaningful transformations. When using neural networks for image classification, these geometric or algebraic invariances can be hard-coded or learned from augmented training data. However, data augmentation strategies are usually limited to one fixed policy. In this talk, we motivate the idea of input-dependent data augmentation, where a separate neural network is trained to transform training images conditional on their features. We also evaluate the model on several classification datasets.
TAG-DS is a hybrid seminar and will be available in-person at the UW Mathematics Department as well as online on Zoom. You can find the link to the Zoom meeting here. If you would like to be added to our mailing list, you can do so by visiting this page.