<content>

  • learning from multiple task
    • transfer learning 
    • multi-task learning

* transfer learning์— ๋Œ€ํ•œ ๋‚ด์šฉ 

hyoeun-log.tistory.com/entry/WEEK5-Transfer-Learning?category=848132

 


Multi-Task Learning

 

1) Multi-Task Learning์ด๋ž€?

  • ์„œ๋กœ ์—ฐ๊ด€์žˆ๋Š” ๊ณผ์ œ๋ฅผ ๋™์‹œ์— ํ•™์Šตํ•จ์œผ๋กœ์จ ๋ชจ๋“  ๊ณผ์ œ ์ˆ˜ํ–‰์˜ ์„ฑ๋Šฅ์„ ์ „๋ฐ˜์ ์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ํ•™์Šต ๋ฐฉ๋ฒ•
  • ๊ด€๋ จ ์žˆ๋Š” ์ž‘์—…๋“ค์˜ ํ‘œํ˜„์„ ๊ณต์œ ํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์ ‘๊ทผ ๋ฐฉ์‹

 


2) Transfer Learning ๋‚ด Multi-task learning

  • Transfer Learning
    • Same Task (transductive transfer learning)
      • domain adaptation
      • cross-lingual learning
    • Different Task (inductive transfer learning)
      • multi-task learning
      • sequential transfer learning

3) Multi-Task learning์˜ ์žฅ์ ๊ณผ ๋‹จ์ 

  • ์žฅ์ 
    • knowledge transfer
      • ์„œ๋กœ ๋‹ค๋ฅธ task๋ฅผ ํ•™์Šตํ•˜๋ฉด์„œ ์–ป์€ ์œ ์šฉํ•œ ์ •๋ณด๊ฐ€ ๊ณต์œ ๋˜์–ด ์„œ๋กœ์—๊ฒŒ ๋„์›€
    • overfitting ๊ฐ์†Œ
      • ๊ด€๋ จ์žˆ๋Š” ์ž‘์—…๋“ค์˜ ํ‘œํ˜„์„ ๊ณต์œ ํ•˜์—ฌ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์ด ์ฆ๊ฐ€
    • computational efficiency
      • ์—ฌ๋Ÿฌ task๋ฅผ ๋™์‹œ์— ํ•™์Šตํ•˜์—ฌ ๊ณ„์‚ฐ ํšจ์œจ์ 
    • real-word application
      • ํ˜„์‹ค์—์„œ๋Š” ๋‹ค์–‘ํ•œ task๊ฐ€ ํ•œ ๋ฒˆ์— ์š”๊ตฌ๋œ๋‹ค
    • ๋ฐ์ดํ„ฐ ์ฆ๊ฐ€์˜ ํšจ๊ณผ
      • ์—ฌ๋Ÿฌ ๊ฐœ์˜ task * ์—ฌ๋Ÿฌ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ -> ๋ฐ์ดํ„ฐ ์ฆ๊ฐ€์˜ ํšจ๊ณผ
  • ๋‹จ์ 
    • negative transfer
      • ๋ถ€์ ํ•ฉํ•œ ํ‘œํ˜„ ์—ญ์‹œ ๊ณต์œ ๋œ๋‹ค.
    • task balancing์˜ ์–ด๋ ค์›€
      • task๋งˆ๋‹ค ํ•™์Šต ๋‚œ์ด๋„๊ฐ€ ํฌ๊ฒŒ ์ฐจ์ด๋‚˜๋ฉด ์ˆ˜๋ ดํ•˜๊ธฐ ์–ด๋ ต๋‹ค
    • ๊ฐ task์˜ ๋ฐ์ดํ„ฐ์…‹์˜ ๊ฐœ์ˆ˜๊ฐ€ ์œ ์‚ฌํ•˜๊ธฐ๋ฅผ ๊ถŒ์žฅ

4) parameter sharing

  • soft parameter sharing
    • ๊ฐ ์ž‘์—…์— ๋Œ€์‘๋˜๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฐ–๋Š” ๊ณ ์œ ํ•œ ๋ชจ๋ธ์ด ์กด์žฌ
    • ๊ฐ ๋ชจ๋ธ ๋งค๊ฐœ๋ณ€์ˆ˜ ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ •๊ทœํ™”ํ•˜์—ฌ ๊ฐ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ์œ ์‚ฌํ•˜๋„๋ก ์œ ๋„
  • hard parameter sharing
    • ์ž‘์—…๋ณ„ ์ถœ๋ ฅ ๋ ˆ์ด์–ด๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ํžˆ๋“  ๋ ˆ์ด์–ด๋ฅผ ๊ณต์œ 
    • ๋™์‹œ์— ํ•™์Šตํ•˜๋Š” ์ž‘์—…์ด ๋งŽ์„์ˆ˜๋ก ๊ณผ์ ํ•ฉ ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋‚ฎ์•„์ง

 


5) mutli-label VS multi-task

  • multi-label
    • ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ์…‹์— ์—ฌ๋Ÿฌ๊ฐ€์ง€ ๋ผ๋ฒจ์ด ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ
  • multi-task
    • ๊ฐ task๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ training set์œผ๋กœ ๊ตฌ์„ฑ

 

* ์ฐธ๊ณ 

vanche.github.io/MTL/

mapadubak.tistory.com/40

 

 

 


Multi-label learning

 

1) Multi-label learning์ด๋ž€?

 

  • ๋ฐ์ดํ„ฐ์…‹์˜ label์ด multi-label์ธ ๊ฒฝ์šฐ
    • multi-class VS multi-label
      • multi-class๋Š” ์—ฌ๋Ÿฌ๊ฐ€์ง€์˜ ํด๋ž˜์Šค ์ค‘ ํ•˜๋‚˜์—๋งŒ ์†ํ•˜๋Š” ๋ฌธ์ œ (์ƒํ˜ธ๋ฐฐํƒ€์ )
        ex) ์ดˆ๋“ฑํ•™์ƒ, ์ค‘ํ•™์ƒ, ๊ณ ๋“ฑํ•™์ƒ
      • multi-label์€ ์—ฌ๋Ÿฌ๊ฐ€์ง€์˜ ํด๋ž˜์Šค ์ค‘ 1๊ฐœ ์ด์ƒ์— ์†ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์ œ
        ex) ์˜ํ™” ์žฅ๋ฅด (๋ฉœ๋กœ, ๋“œ๋ผ๋งˆ, ์ฝ”๋ฏธ๋””, ๋ฒ”์ฃ„, ์Šค๋ฆด๋Ÿฌ, ํŒํƒ€์ง€, ์• ๋‹ˆ๋ฉ”์ด์…˜ ๋“ฑ)
  • ์˜ˆ) ์ž์œจ์ฃผํ–‰
    • ๋ณดํ–‰์ž ์‹๋ณ„ + ์ฐจ๋Ÿ‰ ์‹๋ณ„ + ํ‘œ์ง€ํŒ ์‹๋ณ„ + ์‹ ํ˜ธ๋“ฑ ์‹๋ณ„ + ...
    • Y = [๋ณดํ–‰์ž์œ ๋ฎค ์ž๋™์ฐจ์œ ๋ฌด ํ‘œ์ง€ํŒ์œ ๋ฌด ...]

 

2) ํ•™์Šต ๋ฐฉ๋ฒ•

  • ํ•˜๋‚˜์˜ ๋„คํŠธ์›Œํฌ๊ฐ€ multi label ํ•™์Šต (output node๊ฐ€ ์—ฌ๋Ÿฌ๊ฐœ)
  • multi-label ๊ฐ๊ฐ์˜ ๊ฐœ๋ณ„ ์‹ ๊ฒฝ๋ง ํ•™์Šต

 

'๐Ÿ™‚ > Coursera_DL' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

WEEK5 : CNN (convolutional neural network)  (0) 2020.12.21
WEEK5 : end to end DL  (0) 2020.12.21
WEEK5 : Transfer Learning  (0) 2020.12.20
WEEK5 : traning set๊ณผ dev/test set์˜ ๋ถˆ์ผ์น˜  (0) 2020.12.20
WEEK5 : error analysis (์—๋Ÿฌ ๋ถ„์„)  (0) 2020.12.20

+ Recent posts