Slide 7
Slide 7 text
執筆時点 (2020 年 8 ⽉頃) から過去6ヶ⽉の間に,効率性に焦点を当てた新しいモデルが約 12 種類提案されている.
本稿では,改善観点ごとにトランスフォーマーがまとめられている.
• Axial Transformer (Ho et al., 2019)
• Big Bird (Zaheer et al., 2020)
• Compressive Transformer (Rae et al., 2018)
• ETC (Ainslie et al., 2020)
• Image Transformer (Parmar et al., 2018)
• Linear Transformers (Katharopoulos et al., 2020)
• Linformer (Wang et al., 2020b)
• Longformer (Beltagy et al., 2020)
• Memory Compressed (Liu et al., 2018)
• Performer (Choromanski et al., 2020)
• Reformer (Choromanski et al., 2020)
• Routing Transformer (Roy et al., 2020)
• Set Transformer (Lee et al., 2019)
• Sinkhorn Transformer (Tay et al., 2020b)
• Sparse Transformer (Child et al., 2019)
• Synthesizer (Tay et al., 2020a)
• Transformer-XL (Dai et al., 2019)