<content>

  • train / dev / test split
  • train / test dataset์˜ ๋ถ„ํฌ
  • bias / variance
  • basic recipe for machine learning

 

1. Train / Dev / Test sets

  • ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐ ์žˆ์–ด ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์„ train / dev(validation) / test dataset์œผ๋กœ ๋ถ„๋ฆฌํ•œ๋‹ค 

https://3months.tistory.com/118

    • train dataset
      • ํ•™์Šตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ
    • dev dataset
      • ํ•™์Šต ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” bias๊ฐ€ ์กด์žฌํ•˜๋Š” ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•œ๋‹ค.
    • test dataset
      • ํ•™์Šต ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋‚˜ bias๊ฐ€ ์กด์žฌํ•˜์ง€ ์•Š๋Š” ํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•ด์ค€๋‹ค
  • ์ „ํ†ต์ ์œผ๋กœ train : dev : test = 0.6 : 0.2 : 0.2 ์ •๋„์˜ ๋น„์œจ์œผ๋กœ ๋ถ„ํ• ํ•˜์ง€๋งŒ,
  • ๋ฐ์ดํ„ฐ์…‹์˜ ํฌ๊ธฐ๊ฐ€ ํฌ๋‹ค๋ฉด ์ด๋Ÿฌํ•œ ๋น„์œจ๋กœ ์„ค์ •ํ•  ํ•„์š”๋Š” ์—†๋‹ค.
  • valid / test ๋ชจ๋‘ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ๋ชฉ์ ์ด๋ฏ€๋กœ ๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ์€ ๊ฒฝ์šฐ 0.98 : 0.01 : 0.01์œผ๋กœ ๋‚˜๋ˆ ๋„ ๋œ๋‹ค.

 

 

 


 

2. train / test dataset์˜ ๋ถ„ํฌ

  • train dataset๊ณผ test dataset์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅธ ๊ฒƒ์€ ๊ดœ์ฐฎ๋‹ค.
  • ํ•˜์ง€๋งŒ validation set๊ณผ test set์˜ ๋ถ„ํฌ๋Š” ๋™์ผํ•ด์•ผ ํ•œ๋‹ค.

 


3. bias / variance

http://scott.fortmann-roe.com

 

  • โญ optimal model ํ˜น์€ base model๊ณผ์˜ train cost์˜ ์ฐจ๊ฐ€ ํฐ ๊ฒฝ์šฐ bias๊ฐ€ ํฌ๋‹ค๊ณ  ๋งํ•˜๊ณ  (under-fitting),
  • โญ train cost์™€ valid cost์˜ ์ฐจ๊ฐ€ ํฐ ๊ฒฝ์šฐ variance๊ฐ€ ํฌ๋‹ค๊ณ  ๋งํ•œ๋‹ค (over-fitting).
  • ์šฐ๋ฆฌ์˜ ๋ชฉ์ ์€ bias๋„ ์ค„์ด๊ณ , variance๋„ ์ค„์ด๋Š” ๊ฒƒ์ด๋‹ค!
  • ๊ณผ๊ฑฐ์—๋Š” bias์™€ variance ์‚ฌ์ด trade-off ๊ด€๊ณ„๊ฐ€ ์กด์žฌํ•œ๋‹ค๊ณ  ํ•˜์˜€์œผ๋‚˜,๋”ฅ๋Ÿฌ๋‹์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๊ด€๊ณ„๊ฐ€ ๊นจ์ง€๊ณ  ์žˆ๋‹ค. 
    ๊นŠ์€ layer์™€ ๋งŽ์€ ๋ฐ์ดํ„ฐ๊ฐ€ ์กด์žฌํ•œ๋‹ค๋ฉด bias์™€ variance ๋ชจ๋‘ ๋‚ฎ์ถœ ์ˆ˜ ์žˆ๋‹ค๊ณ  ํ•œ๋‹ค.
    ์ด๊ฒƒ์ด ๋”ฅ๋Ÿฌ๋‹์ด ์‚ฌ๋ž‘๋ฐ›๋Š” ์ด์œ  ์ค‘ ํ•˜๋‚˜๊ฐ€ ์•„๋‹๊นŒ

 


4. basic recipe for machine learning

  • ML ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ๋ณธ์ ์ธ ํ‹€์ด ์žˆ๋‹ค. 
  • 1. bias๊ฐ€ ๋†’์€๊ฐ€?(=underfitting)
    • bias๋ฅผ ์ค„์ด๋Š” ๋ฐ ์ง‘์ค‘ํ•œ๋‹ค. 
    • ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ณ / layer ์ฆ๊ฐ€ / ๋” ๋งŽ์€ iteration ์„ ์‹œ๋„ํ•ด๋ณด๋ผ.
  • 2. bias๋Š” ์ž‘์ง€๋งŒ, variance๊ฐ€ ํฐ ๊ฐ€?(=overfitting)
    • ์ด์ œ๋Š” variance๋ฅผ ์ค„์ด๋Š” ๋ฐ ์ง‘์ค‘ํ•œ๋‹ค.
    • ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ณ / ๋ฐ์ดํ„ฐ ์ถ”๊ฐ€ / ์ •๊ทœํ™”๋ฅผ ์‹œ๋„ํ•ด๋ณด๋ผ.

 

+ Recent posts