SparseBERT: Rethinking the Importance Analysis in Self-attention

     Han Shi 1 Jiahui Gao 2 Xiaozhe Ren 3 Hang Xu 3 Xiaodan Liang 4 Zhenguo Li 3 James T. Kwok 1

              Abstract               include the BERT (Devlin et al., 2019), which achieves state-
    Transformer-based models are popularly used in      of-the-art performance on a wide range of NLP tasks, and
    natural language processing (NLP). Its core com-     GPT-3 (Brown et al., 2020) which applies the Transformer’s
    pone ...