Introduces Cutout, a regularization technique that randomly masks square regions of input images during training. Inspired by dropout but applied to inputs, encouraging models to learn from partially visible objects.