What inductive bias does mixup introduce?

Question: Standard mixup samples two examples and forms convex combinations of both their inputs and labels. What behavior does this most directly encourage? A) Approximately linear predictive behavior between training examples B) Invariance to every transformation preserving the class label C) Maximum-margin separation in the original feature space D) Sparse gradients for examples near the decision boundary Correct: A Explanation: Mixup trains the model on interpolated examples and interpolated targets. It therefore discourages abrupt changes between observed samples and favors simpler, approximately linear transitions in those regions. Topic: advanced ML / regularization / mixup