1. training set๊ณผ test set์ด ๋‹ค๋ฅธ ๊ฒฝ์šฐ

 

  • ๋‘ ๊ฐ€์ง€์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ–ˆ๋‹ค.
    • ์›นํŽ˜์ด์ง€์˜ ๋ฐ์ดํ„ฐ 200,000์žฅ (๊ณ ํ™”์งˆ, ํ”„๋กœํŽ˜์…”๋„ํ•œ ์‚ฌ์ง„)
    • ๋ชจ๋ฐ”์ผ์•ฑ์˜ ๋ฐ์ดํ„ฐ 10,000์žฅ (์ €ํ™”์งˆ, ์•„๋งˆ์ถ”์–ด์˜ ์‚ฌ์ง„)
    • ์‹ค์ œ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ ์ž ํ•˜๋Š” ๋Œ€์ƒ์€ ๋ชจ๋ฐ”์ผ ์•ฑ์˜ ์•„๋งˆ์ถ”์–ด ์‚ฌ์ง„๊ฐ€๋ถ„๋“ค์ด๋‹ค.
  • ์ด๋Ÿฌํ•œ ๊ฒฝ์šฐ train / test set์˜ ๊ตฌ์ถ•?
    • ๋ฐฉ๋ฒ•1 : randomly shuffle & split
      • ์žฅ์  : train / test ๋ชจ๋‘ ๋™์ผํ•œ ๋ถ„ํฌ๋ฅผ ์ง€๋‹Œ๋‹ค.
      • ๋‹จ์  : dev๊ฐ€ ๊ด€์‹ฌ์žˆ๋Š” app์—์„œ์˜ ์ด๋ฏธ์ง€๋ณด๋‹ค ์›นํŽ˜์ด์ง€ ์ด๋ฏธ์ง€์˜ ๋น„์ค‘์ด ๋†’๋‹ค. ๐Ÿ™
      • ์ถ”์ฒœํ•˜์ง€ ์•Š๋Š” ๋ฐฉ๋ฒ•
    • ๋ฐฉ๋ฒ•2 : training์€ web image + mobile ์ผ๋ถ€ / test๋Š” mobile 
      • ์žฅ์  : ์‹ค์ œ ๋™์ž‘ํ•˜๊ธธ ์›ํ•˜๋Š” ํ™˜๊ฒฝ๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ test set์„ ๊ตฌ์ถ•
      • ๋‹จ์  : training set์˜ ๋ถ„ํฌ์™€ dev/test set์˜ ๋ถ„ํฌ์™€ ๋‹ค๋ฅด๋‹ค.
      • ์ด๋ ‡๊ฒŒ ๊ตฌ์„ฑํ•˜๊ธฐ๋ฅผ ์ถ”์ฒœ ๐Ÿ™‚

 


2. train / test ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅผ ๋•Œ bias/variance

 

  • ์—๋Ÿฌ์œจ
    • bayes optimal error = 0%
    • training error = 1%
    • dev error = 10%
  • training set๊ณผ dev set์˜ ์ฐจ์ด๋Š” 1๊ฐœ๊ฐ€ ์•„๋‹ˆ๋‹ค! โญ
    • 2๊ฐ€์ง€๊ฐ€ ๋™์‹œ์— ๋ฐ”๋€Œ์—ˆ๋‹ค
    • 1) training set์€ ๋ณด์•˜์ง€๋งŒ, dev set์€ ํ•™์Šตํ•  ๋•Œ ๋ณด์ง€ ๋ชป ํ–ˆ๋‹ค
    • 2) training set๊ณผ dev set์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅด๋‹ค. 
  • ๋ฌธ์ œ์˜ ์›์ธ (training set๊ณผ dev set์˜ ์ฐจ์ด์ ์ด 1๊ฐœ๊ฐ€ ์•„๋‹ˆ๊ธฐ ๋•Œ๋ฌธ์— ์›์ธ์ด 1๊ฐœ๊ฐ€ ์•„๋‹ˆ๋‹ค) 
    • ์›์ธ1. training set์— over-fitting (variance)
    • ์›์ธ2. train์˜ ๋ถ„ํฌ์™€ dev์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅด๋‹ค (data-mistach)
    • training error์™€ dev error ์‚ฌ์ด์˜ ํฐ ์ฐจ์ด๊ฐ€ ๋‚˜ํƒ€๋‚˜๋Š” ์›์ธ์„  ํŒŒ์•…ํ•˜๊ธฐ ํž˜๋“ค๋‹ค.
      • training-dev set์„ ๋งŒ๋“ค์ž!
  • training-dev set
    • โญ tranining set๊ณผ ๋™์ผํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๋Š” ๋ฐ์ดํ„ฐ์…‹์ด์ง€๋งŒ, ํ•™์Šต์—๋Š” ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š” ๋ฐ์ดํ„ฐ

 

 

์˜ˆ1)

  • training error = 1%
  • training-dev error = 9%
  • dev error = 10%
  • -> variance๊ฐ€ ๋ฌธ์ œ์ด๋‹ค.

์˜ˆ2)

  • training error = 1%
  • training-dev error = 1.5%
  • dev error = 10%
  • -> train / dev ์‚ฌ์ด ๋‹ค๋ฅธ ๋ถ„ํฌ๊ฐ€ ๋ฌธ์ œ์ด๋‹ค (=data mismatch)

์˜ˆ3)

  • human error = 0%
  • training error = 10%
  • training-dev error = 11%
  • dev error = 12%
  • -> avoidable bias๊ฐ€ ๋ฌธ์ œ์ด๋‹ค

์˜ˆ4)

  • human error = 0%
  • training error = 10%
  • training-dev error = 11%
  • dev error = 20%
  • -> bias ๋ฌธ์ œ + train / dev ์‚ฌ์ด ๋‹ค๋ฅธ ๋ถ„ํฌ๊ฐ€ ๋ฌธ์ œ

3. data mismatch ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃจ๋Š” ๋ฒ•

  • ์—๋Ÿฌ ๋ถ„์„ ์ˆ˜ํ–‰
    • training set๊ณผ dev/test set ์ฐจ์ด์ ์— ๋Œ€ํ•ด์„œ ์ดํ•ดํ•œ๋‹ค.
  • dev/test set๊ณผ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ์ˆ˜์ง‘ํ•˜์—ฌ training set์— ์ถ”๊ฐ€ํ•œ๋‹ค.
    • ์œ„์—์„œ ์ˆ˜ํ–‰ํ•œ ์—๋Ÿฌ ๋ถ„์„ ๋‚ด์šฉ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹ ์ถ”๊ฐ€ํ•ด๋ณด๊ธฐ
  • dev/test set๊ณผ ์œ ์‚ฌํ•˜๋„๋ก ๋ฐ์ดํ„ฐ ํ•ฉ์„ฑ ์ˆ˜ํ–‰

+ Recent posts