๐Ÿ“š Generative Adversarial Network ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

์ •๋˜์น˜ยท2022๋…„ 9์›” 7์ผ
0

๋…ผ๋ฌธ๋ฆฌ๋ทฐ

๋ชฉ๋ก ๋ณด๊ธฐ
1/5

0. Abstract


  • Generator์™€ Discriminato ๋‘ ๋ชจ๋ธ์ด ๊ฒฝ์Ÿํ•˜๋Š” ๊ณผ์ •์„ ํ†ตํ•ด ๋‘˜ ๋ชจ๋‘๋ฅผ ๋™์‹œ์— ์ตœ์ ํ™”
  • Generator(์ƒ์„ฑ๋ชจ๋ธ) : ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ํ•™์Šตํ•˜๋Š” ๋ชจ๋ธ
  • Discriminator(ํŒ๋ณ„๋ชจ๋ธ) : ์ƒ์„ฑ๋ชจ๋ธ G๋กœ๋ถ€ํ„ฐ๊ฐ€ ์•„๋‹Œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์—์„œ ์ƒ˜ํ”Œ์ด ๋‚˜์™”์„ ํ™•๋ฅ ์„ ์ถ”์ •ํ•˜๋Š” ๋ชจ๋ธ


1. Introduction


GAN์˜ ํ•ต์‹ฌ ์ปจ์…‰์€ ๊ฐ๊ฐ์˜ ์—ญํ• ์„ ๊ฐ€์ง„ ๋‘ ๋ชจ๋ธ์„ ํ†ตํ•ด ์ ๋Œ€์  ํ•™์Šต์„ ํ•˜๋ฉด์„œ โ€˜์ง„์งœ๊ฐ™์€ ๊ฐ€์งœโ€™๋ฅผ ์ƒ์„ฑํ•ด๋‚ด๋Š” ๋Šฅ๋ ฅ์„ ํ‚ค์›Œ์ฃผ๋Š” ๊ฒƒ์ด๋‹ค.


2. Adversarial Nets


minGmaxDV(D,G)=Exโˆผpdata(x)[logD(x)]+Ezโˆผpz(z)[log(1โˆ’D(G(z)))]min_{_G}max_{_D} V(D, G) = E_{x\sim p_{data}(x)}[logD(x)] + E_{z\sim p_{z}(z)}[log(1-D(G(z)))]

  • ์ฒซ๋ฒˆ์งธ ํ•ญ : ์‹ค์ œ ๋ฐ์ดํ„ฐ x๋ฅผ discriminator ๋ชจ๋ธ์— ๋„ฃ์—ˆ์„ ๋•Œ ๋‚˜์˜ค๋Š” ๊ฒฐ๊ณผ(x๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์—์„œ ๋‚˜์™”์„ ํ™•๋ฅ )์— log๋ฅผ ์ทจํ•ด ์–ป๋Š” ๊ธฐ๋Œ“๊ฐ’
  • ๋‘๋ฒˆ์งธ ํ•ญ : fake ๋ฐ์ดํ„ฐ z๋ฅผ generator ๋ชจ๋ธ์— ๋„ฃ์€ ๊ฒฐ๊ณผ๋ฅผ discriminator ๋ชจ๋ธ์— ๋„ฃ์—ˆ์„๋•Œ ๊ฒฐ๊ณผ์— log(1-๊ฒฐ๊ณผ)๋ฅผ ์ทจํ•ด ์–ป๋Š” ๊ธฐ๋Œ“๊ฐ’

Discriminator D์˜ ์ž…์žฅ์—์„œ G๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ€์งœ ๋ฐ์ดํ„ฐ๊ฐ€ ๋“ค์–ด์˜ค๋ฉด 0์„ ์ถœ๋ ฅํ•˜๊ณ  ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋“ค์–ด์˜ค๋ฉด 1์„ ์ถœ๋ ฅํ•ด์•ผํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ๋ชฉ์ ํ•จ์ˆ˜์˜ ๊ฐ ํ•ญ์ด 0์˜ ๊ฐ’์ด ๋˜๋„๋ก ๋งŒ๋“ค์–ด์•ผ ํ•˜๊ณ , V(D,G)๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฐ’์€ 0์ด ๋œ๋‹ค.

Generator G์˜ ์ž…์žฅ์—์„œ๋Š” D๊ฐ€ ์ƒ˜ํ”Œ์ด ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์—์„œ ๋‚˜์˜จ ๊ฒƒ์œผ๋กœ ํŒ๋‹จํ•˜๊ฒŒ ๋งŒ๋“ค์–ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— D(G(z))์˜ ๊ฐ’์ด 1์ด ๋˜๋„๋ก ๋งŒ๋“ค์–ด์•ผํ•œ๋‹ค. ์ฆ‰, V(D,G)๊ฐ€ ์Œ์˜ ๋ฌดํ•œ๋Œ€ ๊ฐ’(log0)์œผ๋กœ ๊ฐ€๋„๋ก ๋งŒ๋“ค์–ด์•ผํ•œ๋‹ค.


ํ•™์Šต์˜ ์ง„ํ–‰๊ณผ์ •

  • D๊ฐ€ ๋ชจ๋ธ๋งํ•˜๋Š” ์กฐ๊ฑด๋ถ€ ํ™•๋ฅ  ๋ถ„ํฌ : ํŒŒ๋ž€์ƒ‰ ์ ์„ 
  • G๊ฐ€ ๋ชจ๋ธ๋งํ•˜๋Š” ์ƒ์„ฑ ๋ถ„ํฌ(pg) : ๋…น์ƒ‰ ์‹ค์„ 
  • ์‹ค์ œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋ถ„ํฌ(px) : ๊ฒ€์€ ์ ์„ 
  • ํ•˜๋‹จ์˜ ์ˆ˜ํ‰์„  : z๊ฐ€ ๊ท ์ผํ•˜๊ฒŒ ์ƒ˜ํ”Œ๋ง๋˜๋Š” ๋„๋ฉ”์ธ
  • ์ƒ๋‹จ์˜ ์ˆ˜ํ‰์„  : x์˜ ๋„๋ฉ”์ธ
  • ์œ„๋กœ ํ–ฅํ•˜๋Š” ํ™”์‚ดํ‘œ :ย x=G(z)์˜ ๋งคํ•‘์„ ํ†ต๊ณผํ•œ ์ƒ˜ํ”Œ๋“ค์ด ์–ด๋–ค์‹์œผ๋กœ non-uniformํ•œย p_g๋ฅผ ๋‚˜ํƒ€๋‚ด๋„๋ก ํ•˜๋Š”์ง€ ๋ณด์—ฌ์คŒ

(a) ํ•™์Šต ์ดˆ๊ธฐ์˜ ๋ถ„ํฌ ์ƒํƒœ

(b) D์˜ ๋ถ„ํฌ๊ฐ€ ๋ถ„๋ช…ํ•˜๊ฒŒ ๋ฐ์ดํ„ฐ๋ฅผ ํŒ๋ณ„

(c) ์–ด๋Š ์ •๋„ D์˜ ํ•™์Šต์ด ์ด๋ฃจ์–ด์ง€๋ฉด, G๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ๋ชจ์‚ฌํ•˜์—ฌ D๊ฐ€ ํŒ๋ณ„ํ•˜๊ธฐ ํž˜๋“ค๊ฒŒ ํ•™์Šต

(d) ์ด ๊ณผ์ •์„ ๋ฐ˜๋ณตํ•˜์—ฌ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ G์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ถ„ํฌ๊ฐ€ ๊ฑฐ์˜ ๋น„์Šทํ•ด์ ธ D๋Š” 1/2์˜ ๊ฐ’์— ๊ฐ€๊นŒ์šด ํ™•๋ฅ ์„ ๋ณด์—ฌ์คŒ

3. Theoretical Results


Adversarial Nets์ด data์˜ ๋ถ„ํฌ๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ณผ์ •

4. Experiments


0๊ฐœ์˜ ๋Œ“๊ธ€