Researchers from the National University of Singapore and Noah's Ark Lab (Huawei) have developed a new training method that slashes memory requirements with no loss in performance. Their method, called CAME, can match state-of-the-art systems like BERT while using over 40% less memory.