Fair Attribute Classification through Latent Space De-biasing [Kor]

Vikram V. Ramaswamy / Fair Attribute Classification through Latent Space De-biasing / CVPR 2021 Oral

English version of this article is available.

1. Problem definition

์ง€๊ธˆ๊นŒ์ง€ ์ˆ˜๋งŽ์€ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ๊ฐœ๋ฐœ๋˜๋ฉด์„œ ์ธ๊ณต์ง€๋Šฅ์˜ ์„ฑ๋Šฅ์€ ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ชจ๋ธ๋“ค ๋Œ€๋ถ€๋ถ„์€ ๋ฐ์ดํ„ฐ์…‹์˜ ์ „๋ฐ˜์ ์ธ ์˜ˆ์ธก ์ •ํ™•๋„์— ์ดˆ์ ์„ ๋‘๊ณ  ๊ฐœ๋ฐœ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์—, ๋ชจ๋ธ์ด ๋ฐ์ดํ„ฐ์…‹ ๋‚ด์˜ ํŠน์ • ์ง‘๋‹จ์— ๋Œ€ํ•ด ๋ถˆ๋ฆฌํ•œ ํŒ๋‹จ์„ ๋‚ด๋ฆด ์—ฌ์ง€๊ฐ€ ์กด์žฌํ•œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์„œ๊ตฌ๊ถŒ ๊ตญ๊ฐ€์—์„œ ๊ฐœ๋ฐœ๋œ ์–ผ๊ตด ์ธ์‹ AI์˜ ๊ฒฝ์šฐ ์•„์‹œ์•„์ธ์˜ ์–ผ๊ตด์„ ๋ฐฑ์ธ์˜ ์–ผ๊ตด๋ณด๋‹ค ๋” ๋ถ€์ •ํ™•ํ•˜๊ฒŒ ํŒ๋ณ„ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด์™€ ๊ฐ™์€ ํ˜„์ƒ์„ ๊ฐ€๋ฆฌ์ผœ '์ธ๊ณต์ง€๋Šฅ์˜ ๊ณต์ •์„ฑ ๋ฌธ์ œ'๋ผ ๋ถ€๋ฅธ๋‹ค. ์•„๋ฌด๋ฆฌ ์ธ๊ณต์ง€๋Šฅ์˜ ์„ฑ๋Šฅ์ด ์ข‹์•„์ง„๋‹ค๊ณ  ํ•ด๋„, ์ธ๊ณต์ง€๋Šฅ์˜ ๊ณต์ •์„ฑ ๋ฌธ์ œ๊ฐ€ ํ•ด๊ฒฐ๋˜์ง€ ์•Š๋Š”๋‹ค๋ฉด ์ธ๊ณต์ง€๋Šฅ ๋ชจ๋ธ์€ ์žฅ์• ์ธ์ด๋‚˜ ๋…ธ์ธ๊ณผ ๊ฐ™์ด ์‚ฌํšŒ์ ์œผ๋กœ ์†Œ์™ธ๋ฐ›๋Š” ์ง‘๋‹จ์— ๋Œ€ํ•ด ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ์‰ฝ๊ฒŒ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๊ณ , ์ด๋Š” ์‹ฌ๊ฐํ•œ ์‚ฌํšŒ ๋ฌธ์ œ๋ฅผ ์ดˆ๋ž˜ํ•  ๊ฒƒ์ด๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ธ๊ณต์ง€๋Šฅ์„ ๋”์šฑ ๊ณต์ •ํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์€ ๋งค์šฐ ์ค‘์š”ํ•œ ์ผ์ด๋ฉฐ, ์ตœ๊ทผ ์ธ๊ณต์ง€๋Šฅ ํ•™๊ณ„์—์„œ๋Š” ์ธ๊ณต์ง€๋Šฅ์˜ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํฌ์ƒํ•˜์ง€ ์•Š์œผ๋ฉด์„œ๋„ ๊ณต์ •์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ํ™œ๋ฐœํ•˜๊ฒŒ ์—ฐ๊ตฌ๊ฐ€ ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค.

๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๊ณต์ •์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์€ ๋‹ค์–‘ํ•œ๋ฐ, ๋…ผ๋ฌธ ์ €์ž๋Š” ์ ๋Œ€์  ์ƒ์„ฑ ์‹ ๊ฒฝ๋ง(GAN)์„ ํ†ตํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•(Data Augmentation)์„ ์‹œ๋„ํ•œ๋‹ค. ์ฆ‰ GAN์„ ์ด์šฉํ•ด ๊ทธ๋Ÿด๋“ฏํ•œ ์ด๋ฏธ์ง€๋“ค์„ ์ƒ์„ฑํ•œ ๋’ค ์ด๋“ค์˜ ์ž ์žฌ ๊ณต๊ฐ„(latent space)์„ ์ˆ˜์ •ํ•จ์œผ๋กœ์จ, ํŠน์ • ์ง‘๋‹จ์— ๋Œ€ํ•œ ํŽธํ–ฅ์„ฑ์ด ์ œ๊ฑฐ๋˜๋„๋ก ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์„ ๋Š˜๋ฆฌ๋Š” ๋ฐฉ์‹์ด๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ ์ด์™€ ๋น„์Šทํ•œ ์—ฐ๊ตฌ๋Š” ์ด์ „์—๋„ ์žˆ์—ˆ์œผ๋‚˜, ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๋”์šฑ ๋ณต์žกํ•ด์ง€๊ณ  ์—ฐ์‚ฐ๋Ÿ‰์ด ๋Š˜์–ด๋‚œ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ์—ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋…ผ๋ฌธ ์ €์ž๋Š” ๋‹จ ํ•˜๋‚˜์˜ GAN์„ ์‚ฌ์šฉํ•˜๋Š”, ๊ฐ„๋‹จํ•˜๊ณ  ํšจ๊ณผ์ ์ธ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•œ๋‹ค.

2. Motivation

(1) De-biasing methods

๋งŽ์€ ๊ฒฝ์šฐ์— ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๋ถˆ๊ณต์ •์„ฑ์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋‚ด์žฌ๋œ ํŽธํ–ฅ์„ฑ์— ์˜ํ•ด ์ƒ๊ฒจ๋‚œ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์„ ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์„ ์“ฐ๊ธฐ๋„ ํ•˜๊ณ , ๋ชจ๋ธ์˜ ํ•™์Šต ๊ณผ์ •์„ ๋ณด์™„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์“ฐ๊ธฐ๋„ ํ•œ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์„ ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์ทจ์•ฝ ์ง‘๋‹จ์„ ๋Œ€์ƒ์œผ๋กœ ์˜ค๋ฒ„์ƒ˜ํ”Œ๋ง์„ ์ ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•, ์ ๋Œ€์  ํ•™์Šต์„ ์ด์šฉํ•˜๋Š” ๋ฐฉ๋ฒ• ๋“ฑ์ด ์žˆ๋‹ค. ๋ชจ๋ธ์˜ ํ•™์Šต ๊ณผ์ •์„ ๋ณด์™„ํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ๋ชจ๋ธ์˜ ์†์‹คํ•จ์ˆ˜(loss function)์— ๊ณต์ •์„ฑ๊ณผ ๊ด€๋ จ๋œ ๊ทœ์ œ(regularization) ํ•ญ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ• ๋“ฑ์ด ์žˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๊ณต์ •์„ฑ ํ–ฅ์ƒ์„ ์œ„ํ•ด ํ›ˆ๋ จ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์„ ์ค„์ด๋Š” ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•œ๋‹ค.

(2) Generative Adversarial Network (GAN)

์ ๋Œ€์  ์ƒ์„ฑ ์‹ ๊ฒฝ๋ง(GAN)์€ ์ƒ์„ฑ์ž์™€ ํŒ๋ณ„์ž๋กœ ์ด๋ฃจ์–ด์ง„ ์‹ ๊ฒฝ๋ง์ธ๋ฐ, ์—ฌ๊ธฐ์„œ ์ƒ์„ฑ์ž์˜ ํ•™์Šต ๋ฐฉ์‹๊ณผ ํŒ๋ณ„์ž์˜ ํ•™์Šต ๋ฐฉ์‹์€ ์ ๋Œ€์ ์ธ ๊ด€๊ณ„์— ์žˆ๋‹ค. ์ฆ‰ ์ƒ์„ฑ์ž๋Š” ์ž๊ธฐ๊ฐ€ ๊ฑฐ์ง“์œผ๋กœ ๋งŒ๋“ค์–ด ๋‚ธ ๋ฐ์ดํ„ฐ๋ฅผ ํŒ๋ณ„์ž๊ฐ€ ๊ฐ€์งœ๋กœ ์ธ์‹ํ•˜์ง€ ๋ชปํ•˜๋„๋ก ํ•™์Šตํ•˜๊ณ , ํŒ๋ณ„์ž๋Š” ์ƒ์„ฑ์ž๊ฐ€ ์ž๊ธฐ๋ฅผ ์†์ด์ง€ ๋ชปํ•˜๋„๋ก ํ•™์Šตํ•œ๋‹ค. ์ด์™€ ๊ฐ™์ด ์ ๋Œ€์ ์ธ ํ•™์Šต์„ ์‹œํ‚ด์œผ๋กœ์จ ์ง„์งœ์ฒ˜๋Ÿผ ๋ณด์ด๋Š” ๊ฐ€์งœ ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“ค์–ด ๋‚ด๋Š” ์‹ ๊ฒฝ๋ง์ด ๋ฐ”๋กœ ์ ๋Œ€์  ์ƒ์„ฑ ์‹ ๊ฒฝ๋ง์ด๋‹ค. ๊ทธ๋™์•ˆ ์ ๋Œ€์  ์ƒ์„ฑ ์‹ ๊ฒฝ๋ง์€ ๋งŽ์€ ๊ฐœ์„ ์„ ๊ฑฐ์ณ ์™”๊ณ , ์ด์ œ๋Š” ํ˜„์‹ค๊ณผ ๊ตฌ๋ถ„ํ•˜๊ธฐ ๋งค์šฐ ์–ด๋ ค์šด ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์„ ์ •๋„๊ฐ€ ๋˜์—ˆ๋‹ค.

(3) Data augmentation through latent-space manipulation

์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€๋ฅผ ๋ณ€ํ˜•ํ•˜๊ธฐ ์œ„ํ•ด GAN์˜ ์ž ์žฌ ๊ณต๊ฐ„์„ ์ด์šฉํ•ด ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์ž ์žฌ ๊ณต๊ฐ„์ด๋ž€ ์ƒ์„ฑ์ž๊ฐ€ ๋žœ๋คํ•˜๊ฒŒ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ด์šฉํ•˜๋Š” ํŠน์„ฑ๋“ค์˜ ๊ณต๊ฐ„์œผ๋กœ, ์ž ์žฌ ๊ณต๊ฐ„์—๋Š” ์ด๋ฏธ์ง€์˜ ๋‹ค์–‘ํ•œ ์†์„ฑ์ด ์••์ถ•๋˜์–ด ์žˆ๋‹ค. ์ž ์žฌ ๊ณต๊ฐ„์„ ์ž˜ ์กฐ์ž‘ํ•œ๋‹ค๋ฉด ์ด๋ฏธ์ง€์— ํŠน์ • ์†์„ฑ(๋จธ๋ฆฌ ์ƒ‰, ์•ˆ๊ฒฝ ์ฐฉ์šฉ ์—ฌ๋ถ€ ๋“ฑ)์„ ๋ถ€์—ฌํ•˜๊ฑฐ๋‚˜ ์ด๋ฅผ ์กฐ์ ˆํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ๋˜ํ•œ ํŠน์ • ์†์„ฑ์— ๋Œ€ํ•ด์„œ๋งŒ ๊ฐ๊ธฐ ๋‹ค๋ฅธ ๊ฐ’์„ ๊ฐ€์ง„ ์ด๋ฏธ์ง€๋“ค์„ ์ƒ์„ฑํ•จ์œผ๋กœ์จ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ํ•ด๋‹น ์†์„ฑ์— ๋Œ€ํ•ด ์–ผ๋งˆ๋‚˜ ๋ถˆ๊ณต์ •์„ฑ์„ ์ง€๋‹ˆ๊ณ  ์žˆ๋Š” ์ง€ ์ธก์ •ํ•ด ๋ณผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด ๊ณผ์ •์—์„œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๋ถˆ๊ณต์ •์„ฑ๊ณผ ๊ฐ€์žฅ ํฌ๊ฒŒ ์—ฐ๊ด€๋˜์–ด ์žˆ๋Š” ์†์„ฑ์„ ์ฐพ์•„๋‚ผ ์ˆ˜๋„ ์žˆ๋‹ค. ์ด์™€ ๊ฐ™์ด GAN์˜ ์ž ์žฌ ๊ณต๊ฐ„์„ ์ ์ž˜ํžˆ ์ด์šฉํ•œ๋‹ค๋ฉด, ์†์„ฑ ํŽธํ–ฅ์„ฑ์ด ํ•ด์†Œ๋˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ฆ๊ฐ•ํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•˜๋‹ค.

Idea

GAN์˜ ์ž ์žฌ ๊ณต๊ฐ„์„ ์กฐ์ž‘ํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์„ ์กฐ์ ˆํ•˜๋Š” ๊ฒƒ์€ ํšจ์œจ์ ์ธ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์ด๋ผ ํ•  ์ˆ˜ ์žˆ๋‹ค. GAN์„ ์ด์šฉํ•˜๋ฉด ์ด๋ฏธ ๊ฐ€์ง„ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ๋„ ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์–ด ๋‚ผ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€์ ์œผ๋กœ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๋ˆ๊ณผ ์‹œ๊ฐ„์„ ๋‚ญ๋น„ํ•  ํ•„์š”๊ฐ€ ์ค„์–ด๋“ ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ์‹์„ ์œ„ํ•ด ๊ธฐ์กด์— ์‚ฌ์šฉ๋˜์—ˆ๋˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋“ค์€ ์—ฐ์‚ฐ๋Ÿ‰์ด๋‚˜ GAN ๋ชจ๋ธ์˜ ๊ตฌ์กฐ์  ๋ณต์žก์„ฑ์˜ ์ธก๋ฉด์—์„œ ๋ถ„๋ช…ํžˆ ๋‹จ์ ์„ ์ง€๋…”๋‹ค. ํŽธํ–ฅ์„ฑ์„ ์ œ๊ฑฐํ•˜๊ณ ์ž ํ•˜๋Š” ์†์„ฑ์ด ์žˆ์„ ๋•Œ๋งˆ๋‹ค ์ƒˆ๋กœ์šด GAN ๋ชจ๋ธ์„ ๋งŒ๋“ค์–ด ํ›ˆ๋ จ์‹œ์ผฐ์œผ๋ฏ€๋กœ, ๊ณ ๋ ค๋˜๋Š” ์†์„ฑ์ด ์—ฌ๋Ÿฌ ๊ฐœ์ผ ๊ฒฝ์šฐ์—๋Š” ์—ฐ์‚ฐ ์‹œ๊ฐ„์ด ๊ธธ์–ด์ง„๋‹ค๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ์—ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  image-to-image translation GAN๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐ์˜ GAN์„ ์ด์šฉํ•˜๋ฏ€๋กœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋ณต์žก๋„๊ฐ€ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ๋ฌธ์ œ๋„ ์žˆ์—ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์ ๋“ค์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋…ผ๋ฌธ ์ €์ž๋Š” ๋ฐ์ดํ„ฐ์…‹ ์ „์ฒด์—์„œ ํ›ˆ๋ จ๋œ ๋‹จ ํ•˜๋‚˜์˜ GAN์„ ์ด์šฉํ•ด ๋ชจ๋“  ์†์„ฑ์˜ ํŽธํ–ฅ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•œ๋‹ค.

3. Method

3-1. De-correlation definition

์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋ฏธ์ง€์˜ ์†์„ฑ๊ณผ ๋ ˆ์ด๋ธ” ๊ฐ„์— ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๋ฅผ ๋‹ค๋ฃฌ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ฏธ๊ตญ์—์„œ๋Š” ์•ผ์™ธ์—์„œ ์„ ๊ธ€๋ผ์Šค๋ฅผ ์“ฐ๊ณ  ๋‹ค๋‹ˆ๋Š” ์‚ฌ๋žŒ์ด ๋ชจ์ž๋„ ๊ฐ™์ด ์ฐฉ์šฉํ•˜๊ณ  ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ, ์•„๋ž˜์˜ ์‚ฌ์ง„์—์„œ์™€ ๊ฐ™์ด, ์„ ๊ธ€๋ผ์Šค๋ฅผ ์“ฐ๋Š” ๊ฒƒ(์†์„ฑ)๊ณผ ๋ชจ์ž์˜ ์ฐฉ์šฉ ์—ฌ๋ถ€(๋ ˆ์ด๋ธ”) ์‚ฌ์ด์— ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์กด์žฌํ•œ๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒํ™ฉ์—์„œ ์•ผ์™ธ ์ด๋ฏธ์ง€๋“ค์„ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ๊ฑฐ์น˜์ง€ ์•Š๊ณ  ๋ฐ”๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•œ๋‹ค๋ฉด, ๋ชจ์ž์˜ ์ฐฉ์šฉ ์—ฌ๋ถ€๋ฅผ ํŒ๋‹จํ•˜๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์€ ์„ ๊ธ€๋ผ์Šค๋ฅผ ์“ด ์‚ฌ๋žŒ๋“ค๋ณด๋‹ค ์„ ๊ธ€๋ผ์Šค๋ฅผ ์“ฐ์ง€ ์•Š์€ ์‚ฌ๋žŒ๋“ค์— ๋Œ€ํ•ด ๋” ๋ถ€์ •ํ™•ํ•œ ์˜ˆ์ธก์„ ๋‚ด ๋†“์„ ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์‚ฌ์ „์— ์†์„ฑ๊ณผ ๋ ˆ์ด๋ธ” ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์ œ๊ฑฐ๋˜๋„๋ก ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ์ž‘์—…์„ ๊ฑฐ์น˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•˜๋‹ค.

๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ๊ฑฐ์ณ ํŽธํ–ฅ์„ฑ์ด ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ์…‹์„ Xaug ์ด๋ผ ํ•˜๊ณ , ๊ณต์ •์„ฑ๊ณผ ๊ด€๋ จํ•ด์„œ ๊ณ ๋ คํ•˜๋Š” ์†์„ฑ์„ a ๋ผ๊ณ  ํ•˜์ž. ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ์ž„์˜์˜ x โˆˆ Xaug ์— ๋Œ€ํ•˜์—ฌ ์˜ˆ์ธกํ•˜๋Š” ๋ ˆ์ด๋ธ” ๊ฐ’์„ t(x)๋ผ ์ •์˜ํ•˜๊ณ  x์˜ ์˜ˆ์ธก ์†์„ฑ๊ฐ’์„ a(x)๋ผ ํ•˜์ž. ๊ฐ€๋Šฅํ•œ ๋ ˆ์ด๋ธ”์€ -1 ๋˜๋Š” 1 ๋ฟ์ด๋ผ๊ณ  ๊ฐ€์ •ํ•˜๊ณ , ์†์„ฑ๊ฐ’์— ๋Œ€ํ•ด์„œ๋„ ๋˜‘๊ฐ™์ด ๊ฐ€์ •ํ•˜์ž. ๊ทธ๋ ‡๋‹ค๋ฉด ํŽธํ–ฅ์„ฑ์„ ์ œ๊ฑฐํ–ˆ์„ ๋•Œ t(x)=1์ผ ํ™•๋ฅ ์€ a(x)์˜ ๊ฐ’๊ณผ ๋ฌด๊ด€ํ•ด์•ผ ํ•˜๋ฉฐ, ์ˆ˜์‹์œผ๋กœ ํ‘œํ˜„ํ•˜๋ฉด ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

3-2. De-correlation key idea

์ด ๋…ผ๋ฌธ์—์„œ๋Š” ํŽธํ–ฅ์„ฑ์ด ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ์˜ˆ์ธก ๋ ˆ์ด๋ธ”์€ ๋™์ผํ•˜๋ฉด์„œ ์˜ˆ์ธก ์†์„ฑ๊ฐ’์€ ์„œ๋กœ ๋ฐ˜๋Œ€์ธ ์ด๋ฏธ์ง€ ์Œ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•œ๋‹ค. GAN ๋ชจ๋ธ์ด ๊ธฐ์กด ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด ํ›ˆ๋ จ์„ ๋งˆ์ณค๋‹ค๊ณ  ๊ฐ€์ •ํ•˜์ž. ์ž ์žฌ ๊ณต๊ฐ„ ๋‚ด์—์„œ ์ž„์˜๋กœ z๋ผ๋Š” ์ ์„ ์„ ํƒํ•˜๋ฉด, GAN ๋ชจ๋ธ์€ ์  z์„ ํŠน์ •ํ•œ ์ด๋ฏธ์ง€๋กœ ๋ณ€ํ™˜ํ•  ๊ฒƒ์ด๋‹ค. ๊ทธ ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด ๋ถ„๋ฅ˜๊ธฐ ๋ชจ๋ธ์ด ์˜ˆ์ธกํ•˜๋Š” ๋ ˆ์ด๋ธ”์„ t(z)๋ผ ํ•˜๊ณ  ์˜ˆ์ธก ์†์„ฑ๊ฐ’์„ a(z)๋ผ๊ณ  ํ•˜์ž. ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋•Œ ์•„๋ž˜์˜ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋Š” ์ž ์žฌ ๊ณต๊ฐ„ ๋‚ด์˜ ์  zโ€™ ์ƒ์„ฑํ•˜์—ฌ z์™€ ์Œ์„ ์ด๋ฃจ๊ฒŒ ํ•œ๋‹ค.

์ด๋Ÿฐ ์‹์œผ๋กœ ์Œ (z, z')์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ƒ์„ฑํ•œ๋‹ค๋ฉด, ์—์ธก ๋ ˆ์ด๋ธ”์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๊ทธ์— ํ•ด๋‹นํ•˜๋Š” ์ด๋ฏธ์ง€๋“ค์ด ๊ท ๋“ฑํ•œ ์˜ˆ์ธก ์†์„ฑ ๋ถ„ํฌ๋ฅผ ๊ฐ€์งˆ ๊ฒƒ์ด๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ตœ์ข…์ ์œผ๋กœ ์–ป์–ด์ง€๋Š” ๋ฐ์ดํ„ฐ์…‹ Xaug์€ ์†์„ฑ๊ณผ ๋ ˆ์ด๋ธ” ๊ฐ„์˜ ์ƒ๊ด€ ๊ด€๊ณ„๊ฐ€ ํ•ด์†Œ๋˜์—ˆ๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. ์•„๋ž˜์˜ ์‚ฌ์ง„์€ (z, z') ์Œ์„ ์ƒ์„ฑํ•˜๋Š” ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ํ•จ์œผ๋กœ์จ ์†์„ฑ(์„ ๊ธ€๋ผ์Šค ์ฐฉ์šฉ ์—ฌ๋ถ€)๊ณผ ๋ ˆ์ด๋ธ”(๋ชจ์ž ์ฐฉ์šฉ ์—ฌ๋ถ€) ์‚ฌ์ด์˜ ์ƒ๊ด€ ๊ด€๊ณ„๋ฅผ ์ œ๊ฑฐํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค.

3-3. How to calculate zโ€™

๋…ผ๋ฌธ ์ €์ž๋Š” z'์„ ํ•ด์„์ ์œผ๋กœ ๊ตฌํ•˜๊ธฐ ์œ„ํ•˜์—ฌ, ์ž ์žฌ ๊ณต๊ฐ„์ด ์†์„ฑ์— ๋Œ€ํ•ด ์„ ํ˜• ๋ถ„๋ฆฌ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋‹ค(linearly separable)๋Š” ๊ฐ€์ •์„ ๋„์ž…ํ•œ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด ๋‘ ํ•จ์ˆ˜ t(z)์™€ a(z)๋ฅผ ๊ฐ๊ฐ ์ดˆํ‰๋ฉด wt์™€ wa ๋ผ ๊ฐ„์ฃผํ•˜๋Š” ๊ฒƒ์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ์—ฌ๊ธฐ์„œ a(z)์˜ ์ ˆํŽธ์„ ba์ด๋ผ ํ•  ๋•Œ, z'์˜ ์‹์€ ๋…ผ๋ฌธ์— ์˜ํ•˜๋ฉด ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

4. Experiment & Result

Experimental setup

Dataset

ํ•ด๋‹น ๋…ผ๋ฌธ์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ '์„ฑ๋ณ„'์— ๋”ฐ๋ฅธ ๊ณต์ •์„ฑ์„ ์ธก์ •ํ•˜๋Š” ์‹คํ—˜์„ ํ•œ๋‹ค. ์ฆ‰ ์„ฑ๋ณ„์„ ์ œ์™ธํ•œ ์†์„ฑ๋“ค์˜ ๊ฐ’์„ ์˜ˆ์ธกํ•  ๋•Œ, ์˜ˆ์ธก ๊ฒฐ๊ณผ๊ฐ€ ์„ฑ๋ณ„์— ๋”ฐ๋ผ ์–ผ๋งˆ๋‚˜ ์ฐจ์ด๋ฅผ ๋ณด์ด๋Š” ์ง€ ์ธก์ •ํ•œ๋‹ค. ์ €์ž๋Š” ์‹คํ—˜์„ ์œ„ํ•ด, ์œ ๋ช…์ธ์˜ ์–ผ๊ตด ์‚ฌ์ง„์œผ๋กœ ์ด๋ฃจ์–ด์ง„ ๋ฐ์ดํ„ฐ์…‹์ธ CelebA๋ฅผ ์ด์šฉํ•œ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์•ฝ 200๋งŒ ๊ฐœ์˜ ์ด๋ฏธ์ง€๊ฐ€ ๋“ค์–ด ์žˆ๊ณ . ๊ฐ ์ด๋ฏธ์ง€์—๋Š” 40๊ฐœ์˜ ์ด์ง„ ์†์„ฑ(binary attributes)์— ๋Œ€ํ•œ ์ •๋ณด๊ฐ€ ๋‹ด๊ฒจ ์žˆ๋‹ค. ์ €์ž๋Š” 40๊ฐœ์˜ ์†์„ฑ ์ค‘ Male ์†์„ฑ์„ '์„ฑ๋ณ„'๋กœ ๊ฐ„์ฃผํ•˜๋ฉฐ, Male์„ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ 39๊ฐœ์˜ ์†์„ฑ์€ ๊ณต์ •์„ฑ ์ธก์ • ๋‹จ๊ณ„์—์„œ ๋ ˆ์ด๋ธ”๋กœ ์ด์šฉํ•œ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” 39๊ฐœ์˜ ์†์„ฑ์„ ๋ฐ์ดํ„ฐ์˜ ์ผ๊ด€์„ฑ ๋ฐ ์„ฑ๋ณ„๊ณผ์˜ ์—ฐ๊ด€์„ฑ์— ๋”ฐ๋ผ ์•„๋ž˜์˜ ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•œ๋‹ค.

(1) Inconsistently Labeled : ์†์„ฑ๊ฐ’๊ณผ ์‹ค์ œ ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ตํ–ˆ์„ ๋•Œ ์ผ๊ด€์„ฑ์ด ๋ถ€์กฑํ•œ ๊ฒฝ์šฐ

(2) Gender-dependent : ์†์„ฑ๊ฐ’๊ณผ ์‹ค์ œ ์ด๋ฏธ์ง€ ๊ฐ„์˜ ๊ด€๊ณ„๊ฐ€ Male ์—ฌ๋ถ€์— ์˜ํ–ฅ์„ ๋ฐ›๋Š” ๊ฒฝ์šฐ

(3) Geneder-independent : ๊ทธ ์™ธ์˜ ๊ฒฝ์šฐ

Baseline model

์‹คํ—˜์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ค€ ๋ชจ๋ธ(baseline model)๋กœ์„œ ์‚ฌ์ „์— ImageNet์—์„œ ํ›ˆ๋ จ๋œ ResNet-50 ๋ชจ๋ธ์„ ์ด์šฉํ•œ๋‹ค. ํ•ด๋‹น ๋ชจ๋ธ์—์„œ ์™„์ „์—ฐ๊ฒฐ ๊ณ„์ธต(fully-connected layer)์€ ํฌ๊ธฐ 2,048์˜ ์€๋‹‰์ธต์„ ์‚ฌ์ด์— ๋‘” ์ด์ค‘ ์„ ํ˜• ๋ ˆ์ด์–ด๋กœ ๊ต์ฒด๋˜๋ฉฐ, ๋“œ๋กญ์•„์›ƒ ๋ฐ ReLU๊ฐ€ ๋„์ž…๋œ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ CelebA ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์„ ์ด์šฉํ•˜์—ฌ ์ด ๋ชจ๋ธ์„ 20 ์—ํฌํฌ(epoch)๋™์•ˆ ํ•™์Šต์‹œํ‚จ๋‹ค. ํ•™์Šต๋ฅ ์€ 1e-4์ด๊ณ , ๋ฐฐ์น˜ ์‚ฌ์ด์ฆˆ๋Š” 32์ด๋‹ค. ์†์‹คํ•จ์ˆ˜๋กœ ์ด์ง„ ํฌ๋กœ์Šค ์—”ํŠธ๋กœํ”ผ(binary cross entropy)๊ฐ€ ์‚ฌ์šฉ๋˜๋ฉฐ, ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ๋Š” Adam์„ ์ด์šฉํ•œ๋‹ค.

Data Augmentation

ํŽธํ–ฅ์„ฑ ์ œ๊ฑฐ๋ฅผ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๊ณผ์ •์—์„œ ์ ์ง„์  GAN (Progressive GAN)์„ ์ด์šฉํ•œ๋‹ค. ๋‚ด์žฌ ๊ณต๊ฐ„์€ 512์ฐจ์›์œผ๋กœ ์„ค์ •ํ•˜๋ฉฐ, ์ดˆํ‰๋ฉด t(z)์™€ a(z)๋Š” ์„ ํ˜• ์„œํฌํŠธ ๋ฒกํ„ฐ ๋จธ์‹ (linear SVM)์„ ํ†ตํ•ด ํ•™์Šต์‹œํ‚จ๋‹ค.

์ ์ง„์  GAN์„ ํ•™์Šต์‹œํ‚ฌ ๋•Œ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ์ดํ„ฐ์…‹์€ CelebA ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์ด๋‹ค. ํ•™์Šต์ด ๋๋‚˜๋ฉด ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹ Xaug์„ ์–ป์–ด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ํ•˜๋Š”๋ฐ, ์—ฌ๊ธฐ์—๋Š” 1๋งŒ ๊ฐœ์˜ ์ด๋ฏธ์ง€๊ฐ€ ํฌํ•จ๋œ๋‹ค.

Evaluated model & Training setup

ํ‰๊ฐ€์˜ ๋Œ€์ƒ์ด ๋˜๋Š” ๋ชจ๋ธ์€ ๊ธฐ์ค€ ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ ๊ฒƒ์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์ค€ ๋ชจ๋ธ์ด ์›๋ž˜์˜ ํŽธํ–ฅ๋œ ๋ฐ์ดํ„ฐ์…‹ X ์ƒ์—์„œ ํ›ˆ๋ จ๋˜๋Š” ๊ฒƒ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, ํ‰๊ฐ€ ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ์…‹ X์™€ Xaug์„ ํ•จ๊ป˜ ์ด์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋œ๋‹ค. ํ‰๊ฐ€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ์€ ๊ธฐ์ค€ ๋ชจ๋ธ์˜ ํ›ˆ๋ จ๊ณผ ๋™์ผํ•˜๊ฒŒ ์ด๋ฃจ์–ด์ง„๋‹ค.

Evaluation Metrics

๋…ผ๋ฌธ์—์„œ๋Š” ๋ถ„๋ฅ˜ ๋ชจ๋ธ์˜ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋‹ค์Œ์˜ ๋„ค ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ๊ณต์ •์„ฑ์„ ํ‰๊ฐ€ํ•  ๋•Œ๋Š” AP์„ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ ์„ธ ์ง€ํ‘œ๋ฅผ ์ด์šฉํ•˜๋ฉฐ, ์…‹ ๋ชจ๋‘ 0์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ์ข‹์€ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•œ๋‹ค.

(1) AP (Average Precision) : ์ „๋ฐ˜์ ์ธ ์˜ˆ์ธก ์ •ํ™•๋„์ด๋‹ค.

(2) DEO (Difference in Equality of Opportunity) : ์†์„ฑ๊ฐ’์— ๋”ฐ๋ฅธ ๊ฑฐ์ง“ ์Œ์„ฑ๋ฅ ์˜ ์ฐจ์ด์ด๋‹ค.

(3) BA (Bias Amplification) : ์†์„ฑ๊ฐ’์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๋ ˆ์ด๋ธ”๊ฐ’์„ ์‹ค์ œ ๋ฐ์ดํ„ฐ์— ๋น„ํ•ด ์–ผ๋งˆ๋‚˜ ๋” ์ž์ฃผ ์˜ˆ์ธกํ•˜๋Š” ์ง€ ์ธก์ •ํ•˜๋Š” ์ง€ํ‘œ์ด๋‹ค. ์Œ์ˆ˜๊ฐ’์€ ํŽธํ–ฅ์„ฑ์ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์™€ ๋‹ค๋ฅธ ๋ฐฉํ–ฅ์œผ๋กœ ํ˜•์„ฑ๋˜์–ด ์žˆ์Œ์„ ์•”์‹œํ•œ๋‹ค.

(4) KL : ์†์„ฑ๊ฐ’์— ๋”ฐ๋ฅธ ๋ถ„๋ฅ˜๊ธฐ ์ถœ๋ ฅ ์ ์ˆ˜ ๋ถ„ํฌ ๊ฐ„์˜ KL ๋ฐœ์‚ฐ์ด๋‹ค. KL ๋ฐœ์‚ฐ์˜ ๋น„๋Œ€์นญ์„ฑ์„ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๋ถ„ํฌ์˜ ์ˆœ์„œ๋ฅผ ๋ฐ”๊พธ์–ด์„œ ์–ป์€ KL ๋ฐœ์‚ฐ๊ฐ’์„ ๋”ํ•ด ์ค€๋‹ค.

Result

์•„๋ž˜ ํ‘œ๋Š” ๊ธฐ์กด ๋ชจ๋ธ๊ณผ ๋…ผ๋ฌธ์˜ ๋ชจ๋ธ์„ ๋„ค ๊ฐ€์ง€ ์ง€ํ‘œ(AP, DEO, BA, KL)๋ฅผ ํ†ตํ•ด ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ์ด๋‹ค. ๊ฐ๊ฐ์˜ ์ง€ํ‘œ๋Š” ์„ธ ์†์„ฑ ๊ทธ๋ฃน (Inconsistently Labeled, Gender-dependent, Gender-independent)์— ๋Œ€ํ•ด ํ‰๊ฐ€๋˜๋Š”๋ฐ, ํ‘œ์— ์ ํžŒ ๊ฐ’๋“ค์€ ๊ทธ๋ฃน ๋‚ด ์†์„ฑ ๊ฐ๊ฐ์— ๋Œ€ํ•œ ์ง€ํ‘œ๋ฅผ ํ‰๊ท ํ•œ ๊ฒƒ์ด๋‹ค.

ํ‘œ๋ฅผ ๋ณด๋ฉด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ํ›„ ์„ธ ๊ณต์ •์„ฑ ์ง€ํ‘œ(DEO, BA, KL)๊ฐ€ ๋ชจ๋‘ ์ด์ „๋ณด๋‹ค ๊ฐœ์„ ๋œ ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๋ฐ˜๋ฉด ์ „๋ฐ˜์ ์ธ ์˜ˆ์ธก ์ •ํ™•๋„(AP)๋Š” ๊ฐ์†Œํ•œ ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Š” ๊ณต์ •์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ •ํ™•๋„๋ฅผ ์•ฝ๊ฐ„ ํฌ์ƒํ•œ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์ •ํ™•๋„์˜ ๊ฐ์†Œ ํญ์ด ํฌ์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์—, ๋ชจ๋ธ์˜ ๊ณต์ •์„ฑ์ด ์ค‘์š”ํ•œ ๊ฒฝ์šฐ ์ด ๋…ผ๋ฌธ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•˜๋Š” ๊ฒƒ์€ ๊ดœ์ฐฎ์€ ์‹œ๋„๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค.

5. Conclusion

์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๊ณต์ •์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด GAN ๋ชจ๋ธ์˜ ์ž ์žฌ ๊ณต๊ฐ„์„ ์ด์šฉํ•˜์—ฌ ํŽธํ–ฅ์„ฑ์ด ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•˜๊ณ  ์ด๋ฅผ ์ด์šฉํ•ด ์›๋ž˜์˜ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์„ ์ฆ๊ฐ•ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•˜์˜€๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์‹คํ—˜์„ ํ†ตํ•ด, ์ด ๋ฐฉ๋ฒ•์ด ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํฌ์ƒํ•˜์ง€ ์•Š์œผ๋ฉด์„œ ๊ณต์ •์„ฑ์„ ๋†’์ผ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๋ผ๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ๊ฐœ์ธ์ ์œผ๋กœ, ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ์œ„ํ•ด GAN์„ ์ด์šฉํ•˜๋Š” ๊ฒƒ์€ ๋งค๋ ฅ์ ์ธ ๋ฐฉ๋ฒ•์ด๋ผ ์ƒ๊ฐํ•œ๋‹ค. ์ƒˆ๋กœ์šด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ GAN์„ ํ†ตํ•ด ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ, ์ˆ˜์ž‘์—…์— ๋น„ํ•ด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์— ๋“œ๋Š” ์‹œ๊ฐ„ ๋ฐ ๋น„์šฉ์ด ๋งค์šฐ ์ ๋‹ค. ๋˜ํ•œ GAN์—์„œ ์ƒ์„ฑ๋˜๋Š” ์ด๋ฏธ์ง€๊ฐ€ ์‹ค์ œ ์ด๋ฏธ์ง€์™€ ๋งค์šฐ ๋น„์Šทํ•˜๋ฏ€๋กœ, ๊ณ ์ „์ ์ธ ์˜์ƒ ์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ๋”์šฑ ์ž์—ฐ์Šค๋Ÿฌ์šด ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์–ด๋‚ผ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค. ๋˜ํ•œ, ์ด ๋…ผ๋ฌธ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐฉ๋ฒ•์—์„œ ์˜ค์ง ํ•œ ๊ฐœ์˜ GAN ๋ชจ๋ธ์ด ์ด์šฉ๋˜๋ฏ€๋กœ, ๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•œ ๋ฐฉ๋ฒ•์€ ์‹ค์ œ ๊ตฌํ˜„ ๋‚œ์ด๋„ ์ธก๋ฉด์—์„œ ์ด์ ์ด ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•œ๋‹ค.

Take home message (์˜ค๋Š˜์˜ ๊ตํ›ˆ)

GAN์˜ ๋‚ด์žฌ ๊ณต๊ฐ„์„ ์ด์šฉํ•ด ์†์„ฑ๊ณผ ๋ ˆ์ด๋ธ” ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์ œ๊ฑฐ๋œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ณ , ์ด๋ฅผ ํ†ตํ•ด ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๊ณต์ •์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค.

GAN์„ ์ด์šฉํ•ด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ํ•˜๋Š” ๊ฒƒ์€ ํšจ์œจ์„ฑ ๋ฐ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ ๋ฉด์—์„œ ์žฅ์ ์ด ์žˆ๋‹ค.

๋‹จ ํ•˜๋‚˜์˜ GAN ๋ชจ๋ธ์„ ์ด์šฉํ•œ๋‹ค๋Š” ์ ์€ ์‹ค์ œ ๊ตฌํ˜„์˜ ์ธก๋ฉด์—์„œ ๋งค๋ ฅ์ ์ด๋‹ค.

Author / Reviewer information

Author

๊น€๋Œ€ํ˜ (Kim Daehyeok)

  • KAIST ์ „๊ธฐ๋ฐ์ „์ž๊ณตํ•™๋ถ€, U-AIM ์—ฐ๊ตฌ์‹ค

  • ๊ด€์‹ฌ ๋ถ„์•ผ : ์Œ์„ฑ์ธ์‹ ๋ฐ ๊ณต์ •์„ฑ

  • ์—ฐ๋ฝ ์ด๋ฉ”์ผ : kimshine@kaist.ac.kr

Reviewer

  1. Korean name (English name): Affiliation / Contact information

  2. Korean name (English name): Affiliation / Contact information

  3. ...

Reference & Additional materials

  1. Ramaswamy, Vikram V., Sunnie SY Kim, and Olga Russakovsky. "Fair attribute classification through latent space de-biasing." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

  2. https://github.com/princetonvisualai/gan-debiasing

  3. Rameen Abdal, Yipeng Qin, and Peter Wonka. Im- age2StyleGAN: How to embed images into the StyleGAN latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

  4. Mohsan Alvi, Andrew Zisserman, and Christoffer Nella ฬŠker. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings. In Proceedings of the European Conference on Computer Vision (ECCV), 2018.

  5. Sina Baharlouei, Maher Nouiehed, Ahmad Beirami, and Meisam Razaviyayn. Re ฬnyi fair inference. In Proceedings of the International Conference on Learning Representations (ICLR), 2020.

  6. GuhaBalakrishnan,YuanjunXiong,WeiXia,andPietroPer- ona. Towards causal benchmarking of bias in face analysis algorithms. In Proceedings of European Conference on Com- puter Vision (ECCV), 2020.

  7. DavidBau,Jun-YanZhu,JonasWulff,WilliamPeebles,Hen- drik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a GAN cannot generate. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.

  8. RachelK.E.Bellamy,KuntalDey,MichaelHind,SamuelC. Hoffman, Stephanie Houde, Kalapriya Kannan, Pranay Lo- hia, Jacquelyn Martino, Sameep Mehta, Aleksandra Mo- jsilovic, Seema Nagar, Karthikeyan Natesan Ramamurthy, John Richards, Diptikalyan Saha, Prasanna Sattigeri, Monin- der Singh, Kush R. Varshney, and Yunfeng Zhang. AI Fairness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted algorithmic bias, Oct. 2018.

  9. Steffen Bickel, Michael Bru ฬˆckner, and Tobias Scheffer. Dis- criminative learning under covariate shift. Journal of Machine Learning Research, 10(Sep):2137โ€“2155, 2009.

  10. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer program- mer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, pages 4349โ€“4357, 2016.

  11. Joy Buolamwini and Timnit Gebru. Gender shades: Intersec- tional accuracy disparities in commercial gender classification. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 77โ€“91, 2018.

  12. Toon Calders, Faisal Kamiran, and Mykola Pechenizkiy. Building classifiers with independency constraints. In 2009 IEEE International Conference on Data Mining Workshops, pages 13โ€“18. IEEE, 2009.

  13. Mingliang Chen and Min Wu. Towards threshold invariant fair classification. In Proceedings of the Conference on Un- certainty in Artificial Intelligence (UAI), 2020.

  14. Kristy Choi, Aditya Grover, Rui Shu, and Stefano Ermon. Fair generative modeling via weak supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2020.

  15. Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: Unified genera- tive adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.

  16. Emily Denton, Ben Hutchinson, Margaret Mitchell, and Timnit Gebru. Image counterfactual sensitivity analysis for detecting unintended bias. In CVPR 2019 Workshop on Fair- ness Accountability Transparency and Ethics in Computer Vision, 2019.

  17. Charles Elkan. The foundations of cost-sensitive learning. In Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI), volume 17, pages 973โ€“978. Lawrence Erlbaum Associates Ltd, 2001.

  18. FAIR HDGAN. Pytorch GAN Zoo.

  19. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672โ€“2680, 2014.

  20. Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of Wasserstein GANs. In Advances in Neural Information Pro- cessing Systems, pages 5767โ€“5777, 2017.

  21. Moritz Hardt, Eric Price, and Nati Srebro. Equality of oppor- tunity in supervised learning. In Advances in Neural Informa- tion Processing Systems, pages 3315โ€“3323, 2016.

  22. Bharath Hariharan and Ross Girshick. Low-shot visual recog- nition by shrinking and hallucinating features. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 3018โ€“3027, 2017.

  23. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770โ€“778, 2016.

  24. Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Over- coming bias in captioning models. In Proceedings of Euro- pean Conference on Computer Vision (ECCV), pages 793โ€“ 811. Springer, 2018.

  25. Khari Johnson. Google Cloud AI removes gender labels from Cloud Vision API to avoid bias, 02 2020.

  26. Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Proceedings of the International Conference on Learning Representations (ICLR), 2018.

  27. Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401โ€“4410, 2019.

  28. Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015.

  29. Steven Liu, Tongzhou Wang, David Bau, Jun-Yan Zhu, and Antonio Torralba. Diverse image generation via self- conditioned GANs. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2020.

  30. Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3730โ€“3738, 2015.

  31. Vishnu Suresh Lokhande, Aditya Kumar Akash, Sathya N. Ravi, and Vikas Singh. FairALM: Augmented lagrangian method for training fair models with little regret. In Proceed- ings of European Conference on Computer Vision (ECCV), 2020.

  32. Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classi- fier from biased classifier. In Advances in Neural Information Processing Systems, 2020.

  33. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825โ€“2830, 2011.

  34. Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Com- puter Vision, 115(3):211โ€“252, 2015.

  35. Hee Jung Ryu, Hartwig Adam, and Margaret Mitchell. Inclu- siveFaceNet: Improving face attribute detection with race and gender diversity. In International Conference on Machine Learning (ICML) FATML Workshop, 2018.

  36. Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Advances in Neural Information Pro- cessing Systems, pages 2234โ€“2242, 2016.

  37. PrasannaSattigeri,SamuelCHoffman,VijilChenthamarak- shan, and Kush R Varshney. Fairness GAN: Generating datasets with fairness properties using a generative adver- sarial network. IBM Journal of Research and Development, 63(4/5):3โ€“1, 2019.

  38. ViktoriiaSharmanska,LisaAnneHendricks,TrevorDarrell, and Novi Quadrianto. Contrastive examples for addressing the tyranny of the majority, 2020.

  39. Yujun Shen, Jinjin Gu, Xiaoou Tang, and Bolei Zhou. Inter- preting the latent space of GANs for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243โ€“9252, 2020.

  40. AngelinaWang,ArvindNarayanan,andOlgaRussakovsky. REVISE: A tool for measuring and mitigating bias in visual datasets. In Proceedings of the European Conference on Computer Vision (ECCV), 2020.

  41. Angelina Wang and Olga Russakovsky. Directional bias amplification. arXiv preprint arXiv:2102.12594, 2021.

  42. TianluWang,JieyuZhao,MarkYatskar,Kai-WeiChang,and Vicente Ordonez. Balanced datasets are not enough: Estimat- ing and mitigating gender bias in deep image representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5310โ€“5319, 2019.

  43. ZeyuWang,KlintQinami,IoannisKarakozis,KyleGenova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

  44. Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fair- GAN: Fairness-aware generative adversarial networks. In 2018 IEEE International Conference on Big Data (Big Data), pages 570โ€“575. IEEE, 2018.

  45. Kaiyu Yang, Klint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 547โ€“558, 2020.

  46. Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Ro- driguez, and Krishna P Gummadi. Fairness beyond disparate treatment & disparate impact: Learning classification without disparate mistreatment. In Proceedings of the International Conference on World Wide Web (WWW), pages 1171โ€“1180, 2017.

  47. Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. Mitigating unwanted biases with adversarial learning. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), pages 335โ€“340, 2018.

  48. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gen- der bias amplification using corpus-level constraints. In Pro- ceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2017.

  49. Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. In- domain GAN inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020.

  50. XinyueZhu,YifanLiu,JiahongLi,TaoWan,andZengchang Qin. Emotion classification with data augmentation using generative adversarial networks. In Dinh Phung, Vincent S. Tseng, Geoffrey I. Webb, Bao Ho, Mohadeseh Ganji, and Lida Rashidi, editors, Advances in Knowledge Discovery and Data Mining (KDD), pages 349โ€“360, Cham, 2018. Springer International Publishing.

Last updated