SkipBERT

AbstractIn this paper, we propose SkipBERT to accelerate BERT inference by skipping the computation of shallow layers. To achieve this, our approach encodes small text chunks into independent representations, which …...