The speaker discusses data augmentation for vision, which involves image filtering operations like blurring, Instagram filters, and saturating images. They argue that data augmentation may involve more learning than the learning process itself.