Early Stopping & Data Augmentation — Checkpoint Quiz

A is correct. Early stopping monitors validation loss and halts training (or restores an earlier checkpoint) once that loss stops improving, capturing the model at the point of best generalization rather than letting it keep memorizing training data.

A is correct. Augmentation creates many plausible variations of each training example, so the model can't simply memorize the exact pixels of a given image. It also teaches the model that, e.g., a flipped cat is still a cat.

B is correct. Augmentations should reflect transformations the model should genuinely be invariant to. Flipping a digit or a word horizontally can change its meaning entirely, so applying that augmentation would teach the model an incorrect invariance.