Normalizing self-supervised learning for provably reliable Change Point Detection