“Techniques from computer vision, like distorting or hiding parts of an image to make the learning task harder, could be adapted for text and code data to improve model capabilities.”