Groups
Category
Data augmentation expands the training distribution by applying label-preserving transformations to inputs, which lowers overfitting and improves generalization.
The Information Bottleneck (IB) principle formalizes learning compact representations T that keep only the information about X that is useful for predicting Y.