[Transformer] Time-Series Transformer

2023. 1. 27. 23:13
๐Ÿง‘๐Ÿป‍๐Ÿ’ป์šฉ์–ด ์ •๋ฆฌ
 
Transformer

 

Transformer

  • NLP๋ฅผ ์œ„ํ•ด ๋งŒ๋“ค์–ด์ง„ ๋ชจ๋ธ
  • ํŠน์ • ๋ฌธ์žฅ์ด ๋“ฑ์žฅํ•  ํ™•๋ฅ ์„ ๊ณ„์‚ฐํ•ด์ฃผ๋Š” ๋ชจ๋ธ
  • ์ด๋ฏธ์ง€์™€ ์‹œ๊ณ„์—ด๋กœ๋„ ํ™•์žฅ๋˜์–ด ์‚ฌ์šฉ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
  • Attention์„ ๋ณ‘๋ ฌ์ ์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
  • encoder์™€ decoder ์กด์žฌ
  • embedding vector
  • semantic relationship
    • word imbedding
  • positional encoding
  • feed forward neural network
  • self attention
    • query, key, value
    • ์ž…๋ ฅ ๋ฒกํ„ฐ์— ๋Œ€ํ•ด ์„ธ ๊ฐœ์˜ ๋ฒกํ„ฐ ์ƒ์„ฑ
    • query์— ๋Œ€ํ•œ Key ๊ฐ’
  • multi-headed attention
  • position-wise feed-forward neural network
  • masked multi-head attention
  • Linear Layer
  • Softmax Layer
  • batch normalization vs. layer normalization
  • freeze
  • TST

 

BELATED ARTICLES

more