๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

GAN

CS/์ธ๊ณต์ง€๋Šฅ 2024. 5. 28. Generative Adversarial Network Discriminative vs generative modelsDiscriminative ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค๋“ค ๊ฐ„์˜ ๊ฒฝ๊ณ„๋ฅผ ํ•™์Šตํ•˜๋Š”๋ฐ ์ค‘์ ์„ ๋‘”๋‹ค.์ด๋Š” ํŠน์ง• X๊ฐ€ ์ฃผ์–ด์กŒ์„ ๋•Œ, ๋ ˆ์ด๋ธ” y์˜ ํ™•๋ฅ  P(y|X)์„ ์ง์ ‘ ๋ชจ๋ธ๋งํ•˜์—ฌ ๊ฒฐ์ • ๊ฒฝ๊ณ„๋ฅผ ์ฐพ๋Š” ๊ฒƒ์ด๋‹ค.์ด ๋ชจ๋ธ๋“ค์€ ์˜ˆ์ธก ์˜ค๋ฅ˜๋ฅผ ์ธก์ •ํ•˜๋Š” ๋น„์šฉ ํ•จ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•˜์—ฌ ํ›ˆ๋ จ๋˜๊ณ , ํ›ˆ๋ จ ๊ณผ์ •์—์„œ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์กฐ์ •ํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ตœ์†Œํ™”ํ•œ๋‹ค.ex) ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€, CNN, RNN Generative model ์€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค. ์ƒ์„ฑ ๋ชจ๋ธ์€ ํŠน์ง•๊ณผ ๋ ˆ์ด๋ธ”์˜ ๊ฒฐํ•ฉ ํ™•๋ฅ  ๋ถ„ํฌP(X,y)๋ฅผ ํ•™์Šตํ•˜๋Š”๋ฐ ์ค‘์ ์„ ๋‘”๋‹ค. โ“๊ฒฐํ•ฉ ํ™•๋ฅ  ๋ถ„ํฌ ๐‘ƒ(๐‘‹,๐‘ฆ)ํŠน์ง• ๐‘‹์™€ ๋ ˆ์ด๋ธ” ๐‘ฆ ์˜ ๊ฒฐํ•ฉ ํ™•๋ฅ  ๋ถ„ํฌ ๐‘ƒ(๐‘‹,๐‘ฆ)๋Š” ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ์—์„œ ํŠน์ง•๊ณผ ๋ ˆ์ด๋ธ”์ด ๋™..

ํ‹ฐ์Šคํ† ๋ฆฌํˆด๋ฐ”