Masked Autoencoders (MAE) a self-supervised representation learner that reconstructs missing parts but is not a generator