As far as I understand, in an ideal world if we apply data attribution methods to a model, the model is 1) supposed to be trained from scratch (not pre-trained on a larger dataset); 2) not supposed to involve data augmentation. But how are we supposed to train any (non-MNIST) models like that so that they don't perform bad?
As far as I understand, in an ideal world if we apply data attribution methods to a model, the model is 1) supposed to be trained from scratch (not pre-trained on a larger dataset); 2) not supposed to involve data augmentation. But how are we supposed to train any (non-MNIST) models like that so that they don't perform bad?