#image-augmentation

Question 1: Single choice

You train the same ResNet-18 on CIFAR-10 twice, once with random horizontal flips and once with no augmentation, and compare training and test accuracy. Which result supports the claim that image augmentation mitigates overfitting?

With augmentation, training accuracy rises faster, so the model fits the training images more closely1
Without augmentation, training and test accuracy are both lower than with augmentation2
Without augmentation, training accuracy is higher but test accuracy lower, so the train–test gap is wider3
Both runs reach the same test accuracy, but the run with augmentation finishes its epochs sooner4

Timed test