A curve of accuracy against dataset size, from 100 images to a million. Trained from scratch, a model needs a huge pile of images before it is any good.
Training from scratch, accuracy climbs slowly and needs an enormous amount of data before it becomes useful at all.
