Group Study (2020-2021)/Deep Learning

[DeepSleep] ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ ์Šคํ„ฐ๋”” 6์ฃผ์ฐจ

morijwana 2021. 8. 16. 21:29

๐Ÿ“–  6์ฃผ์ฐจ ๋ฐœํ‘œ ๋‚ด์šฉ

โ„๏ธ ํ˜œ์ฃผ

  • ์„ ์ •ํ•œ ๋…ผ๋ฌธ: Neural Machine Translation by Jointly Learning to Align and Translate [pdf] 
  • ๋ฐœํ‘œ ์ž๋ฃŒ: https://github.com/dsc-sookmyung/2021-DeepSleep-Paper-Review/blob/main/Week6/align.md
  • ์ฃผ์ œ: Alignment model ์˜ ๋“ฑ์žฅ๊ณผ ์ž…๋ ฅ ๋ฌธ์žฅ ๋ฒกํ„ฐ์˜ ์—ฐ๊ด€ ์ˆœ์œ„ ์ฑ…์ •์— ๋”ฐ๋ฅธ ๋ฒˆ์—ญ ํšจ์œจ ํ–ฅ์ƒ
  • ๋ฐฐ๊ฒฝ: ๊ธฐ์กด์ฒ˜๋Ÿผ ์ž…๋ ฅ ๋ฌธ์žฅ์„ ๊ณ ์ •๋œ ๊ธธ์ด์˜ context vector ๋กœ ๋ณ€ํ™˜ํ•  ์‹œ, ๊ธธ์ด๊ฐ€ ๊ธด ์ž…๋ ฅ ๋ฌธ์žฅ์— ๋Œ€ํ•ด์„œ๋Š” ๋ฒˆ์—ญ ์„ฑ๋Šฅ์ด ๊ธ‰๊ฒฉํžˆ ์ €ํ•˜๋˜๋Š” ๋ฌธ์ œ์ ์ด ๋ฐœ์ƒ
  • ๋‚ด์šฉ
    1. decoder ์—์„œ output ์„ ์ถœ๋ ฅํ•  ๋•Œ, ์ž…๋ ฅ ๋ฌธ์žฅ์„ ์ˆœ์ฐจ์ ์œผ๋กœ ํƒ์ƒ‰ํ•ด์„œ ํ˜„์žฌ ์ƒ์„ฑํ•˜๋ ค๋Š” decoder์˜ output ๊ณผ ๊ฐ€์žฅ ๊ด€๋ จ์žˆ๋Š” ์˜์—ญ์„ ์ ์šฉ์‹œํ‚ด
    2. ๋”ฐ๋ผ์„œ ๊ณ ์ •๋œ ๊ธธ์ด์˜ context vector ๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ , encoder ์—์„œ ์ƒ์„ฑํ•œ ์—ฌ๋Ÿฌ context vector ๋ฅผ ๊ณ„์†ํ•ด์„œ ์ฐธ์กฐํ•˜๋ฏ€๋กœ ๋ฌธ์žฅ์˜ ๊ธธ์ด๊ฐ€ ๊ธธ์–ด๋„ ์„ฑ๋Šฅ ์œ ์ง€ ๊ฐ€๋Šฅ

 

โ„๏ธ ๋„์—ฐ

 


๐Ÿ“–  7์ฃผ์ฐจ ๋ฐœํ‘œ ๊ณ„ํš

โ„๏ธ ์ˆ˜์—ฐ

  • ์„ ์ •ํ•œ ๋…ผ๋ฌธ: Attention Is All You Need [pdf]
  • ์ฃผ์ œ: ์ˆœํ™˜ ๊ตฌ์กฐ ์—†์ด Attention๋งŒ์„ ์‚ฌ์šฉํ•œ ์ƒˆ๋กœ์šด ๋ชจ๋ธ(Transformer) ์ œ์‹œ