• Turing Post Korea
  • Posts
  • ๐ŸŒFOD#83: GAN แ„‹แ…ตแ„Œแ…ณ แ„‡แ…ขแ†จ!

๐ŸŒFOD#83: GAN แ„‹แ…ตแ„Œแ…ณ แ„‡แ…ขแ†จ!

แ„‡แ…กแ„ˆแ…ต แ„‹แ…ฎแ†ทแ„Œแ…ตแ†จแ„‹แ…ตแ„‚แ…ณแ†ซ แ„‰แ…ฆแ„‰แ…กแ†ผแ„‹แ…ฆ แ„Œแ…ฅแ†ผแ„‰แ…ตแ†ซแ„‹แ…ฅแ†นแ„‹แ…ณแ†ฏ แ„„แ…ข, แ„Œแ…กแ†ทแ„แ…กแ†ซ แ„„แ…กแ†ซ แ„‰แ…ขแ†ผแ„€แ…กแ†จแ„ƒแ…ฉ แ„€แ…ซแ†ซแ„Žแ…กแ†ญแ„Œแ…ญ.

๋“ค์–ด๊ฐ€๋ฉฐ

์ง€๋‚œ ์ฃผ์˜ ํ—ค๋“œ๋ผ์ธ์€ ์•„๋ฌด๋ž˜๋„ CES ๊ด€๋ จ ์†Œ์‹๋“ค์ด ์ ๋ นํ•˜์ง€ ์•Š์•˜๋‚˜ - AI ๊ด€์ ์—์„œ๋Š”์š” - ์‹ถ์Šต๋‹ˆ๋‹ค. ์•„์ง๊นŒ์ง€๋„ CES ํ›„์† ๊ธฐ์‚ฌ๋ผ๋“ ๊ฐ€ CES์˜ ์˜๋ฏธ, ์•ž์œผ๋กœ์˜ (๊ธ์ •์ , ๋ถ€์ •์ ์ธ) ์ „๋ง ๋“ฑ์ด ๊ณ„์† ์ด์–ด์ง€๊ณ  ์žˆ์œผ๋‹ˆ๊นŒ์š”.

์˜ค๋Š˜ ํŠœ๋ง ํฌ์ŠคํŠธ ์ฝ”๋ฆฌ์•„์—์„œ๋Š”, ์ž ์‹œ ๋ฐ”์˜๊ฒŒ ๋Œ์•„๊ฐ€๋Š” ์„ธ์ƒ์—์„œ ์‹œ์„ ์„ ๋Œ๋ ค์„œ, ๋‹ค์†Œ ๊ณ ์ „์ ์ธ ์ฃผ์ œ๋ฅผ ํ•œ ๋ฒˆ ์‚ดํŽด๋ณผ๊นŒ ํ•ฉ๋‹ˆ๋‹ค - ๋ฐ”๋กœ โ€˜GAN (Generative Adversarial Network; ์ƒ์„ฑํ˜• ์ ๋Œ€์  ์‹ ๊ฒฝ๋ง)โ€™ ์ด์•ผ๊ธฐ์ธ๋ฐ์š”, ์ง€๋‚œ 1์›” 9์ผ ๋ฐœํ‘œ๋œ โ€˜The GAN Is Dead; Long Live the GAN! A Modern GAN Baselineโ€™์ด๋ผ๋Š” ๋…ผ๋ฌธ์„ ๋ณด๊ณ  ํ•œ ๋ฒˆ ์ด์•ผ๊ธฐํ•ด๋ณด๊ณ  ์‹ถ๋‹ค๋Š” ์ƒ๊ฐ์ด ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. โ€˜ํ˜„๋Œ€์  AIโ€™๋ผ๋Š” ๊ด€์ ์—์„œ ํ•œ ํš์„ ๊ทธ์€ ์‹ ๊ฒฝ๋ง ์•„ํ‚คํ…์ฒ˜ ์ค‘ ํ•˜๋‚˜์ž„์— ๋ถ„๋ช…ํ•˜์ง€๋งŒ, ์‹ค์ œ ์ ์šฉ์˜ ๋‚œ์ ๋“ค ๋•Œ๋ฌธ์— ์ง€๊ธˆ์€ ๋””ํ“จ์ „ (Diffusion) ๋ชจ๋ธ ๊ณ„์—ด์—๊ฒŒ ๊ทธ ์ž๋ฆฌ๋ฅผ ๋‚ด์ฃผ์—ˆ๋‹ค๊ณ  ํ•  ์ˆ˜๋„ ์žˆ๋Š” GAN. GAN์€ ๊ณผ์—ฐ ML์—์„œ ๊ฐ€์žฅ ๋งค๋ ฅ์ ์ธ ์•„์ด๋””์–ด ์ค‘ ํ•˜๋‚˜๋ผ๋Š” ํƒ€์ดํ‹€์„ ์•ž์œผ๋กœ๋„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”?

GAN์˜ ํƒ„์ƒ: ๋‘ ๋„คํŠธ์›Œํฌ์˜ ๊ฒŒ์ž„

โ€˜GAN (Generative Adversarial Networks; ์ƒ์„ฑํ˜• ์ ๋Œ€์  ์‹ ๊ฒจ๊ฒฝ๋ง)โ€™ ๋…ผ๋ฌธ์€ 2014๋…„ ์ด์•ˆ ๊ตฟํŽ ๋กœ์šฐ (Ian Goodfellow)์™€ ๊ทธ๊ฐ€ ์ด๋ˆ ํŒ€์ด ์†Œ๊ฐœํ–ˆ์ฃ .

์ด์•ˆ ๊ตฟํŽ ๋กœ์šฐ. Image Credit: ๋‚ด์™ธ๋ฐฉ์†ก

GAN์˜ ๊ฐœ๋…์€ ๋‹จ์ˆœํ•˜๋ฉด์„œ๋„ ํ˜์‹ ์ ์ด์—ˆ์–ด์š”: ์ƒ์„ฑ์ž (Generator)์™€ ํŒ๋ณ„์ž (Discriminator), ๋‘ ๊ฐœ์˜ ์‹ ๊ฒฝ๋ง์ด ์ œ๋กœ์„ฌ ๊ฒŒ์ž„ ์กฐ๊ฑด ์•„๋ž˜์„œ ๊ฒฝ์Ÿํ•˜๋Š” ๊ฒ๋‹ˆ๋‹ค.

(*ํŽธ์ง‘์ž ์ฃผ: ๋ชฌํŠธ๋ฆฌ์˜ฌ ๋Œ€ํ•™๊ต์˜ ์•„๋ก  ์ฟ ๋ฅด๋นŒ ๊ต์ˆ˜์—๊ฒŒ ๋“ค์€ ๋ฐ”๋กœ๋Š”, ์ด์•ˆ ๊ตฟํŽ ๋กœ์šฐ๊ฐ€ ๋™๋ฃŒ๋“ค๊ณผ ๋ชฌํŠธ๋ฆฌ์˜ฌ์˜ ํ•œ ๋ฐ”์—์„œ ์ˆ ์„ ๋งˆ์‹œ๋ฉด์„œ ์ด์•ผ๊ธฐ๋ฅผ ๋‚˜๋ˆ„๋‹ค๊ฐ€ ๊ฐ‘์ž๊ธฐ ์•„์ด๋””์–ด๊ฐ€ ๋– ์˜ฌ๋ผ์„œ, ์ˆ ์ด ์‚ด์ง ์ทจํ•œ ์ƒํƒœ์—์„œ ์—ฐ๊ตฌ์‹ค๋กœ ๋Œ์•„๊ฐ€์„œ ์ฝ”๋”ฉ์„ ํ•˜๋ฉด์„œ GAN์„ ๋งŒ๋“ค๊ฒŒ ๋˜์—ˆ๋‹ค๊ณ  ํ•ด์š” ^.^;)

GAN ์ปจ์…‰

  • ์ƒ์„ฑ์ž (Generator)
    ์ด ๋„คํŠธ์›Œํฌ๋Š” ๋ฌด์ž‘์œ„์˜ ๋…ธ์ด์ฆˆ๋กœ๋ถ€ํ„ฐ ๊ฐ€์งœ ๋ฐ์ดํ„ฐ (์˜ˆ: ์ด๋ฏธ์ง€, ์˜ค๋””์˜ค, ํ…์ŠคํŠธ)๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ์ด ๋„คํŠธ์›Œํฌ์˜ ๋ชฉํ‘œ๋Š” ๋‹ค๋ฅธ ๋„คํŠธ์›Œํฌ (ํŒ๋ณ„์ž)๊ฐ€ ๊ฐ€์งœ์ธ์ง€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์„ ์ •๋„๋กœ ์‹ค์ œ์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฑฐ์ฃ .

  • ํŒ๋ณ„์ž(Discriminator)
    ์ด ๋„คํŠธ์›Œํฌ๋Š” โ€˜์‹ฌํŒโ€™ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ (์‹ค์ œ์™€ ๊ฐ€์งœ ๋ชจ๋‘)๋ฅผ ๋ณด๊ณ  ๊ทธ๊ฒŒ ์ง„์งœ์ธ์ง€ ์ƒ์„ฑ์ž๊ฐ€ ๋งŒ๋“  ๊ฒƒ์ธ์ง€ ํŒ๋‹จํ•˜๋ ค๊ณ  ํ•˜์ฃ .

์ด๋Ÿฐ ๊ตฌ์กฐ๋กœ ์ ๋Œ€์  ํŠธ๋ ˆ์ด๋‹ (Adversarial Training)์„ ๋ฐ˜๋ณตํ•˜๋ฉด์„œ ๋‘ ๋„คํŠธ์›Œํฌ๊ฐ€ ๋ชจ๋‘ ๋ฐœ์ „ํ•˜๊ฒŒ๋” ๋งŒ๋“ค์–ด์„œ, ๊ฒฐ๊ตญ โ€˜์‹ค์ œ์™€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†๋Š”โ€™ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

โ€˜GANโ€™์˜ ๊ฐœ๋…์€ ํฅ๋ฏธ๋กœ์šธ ๋ฟ ์•„๋‹ˆ๋ผ ๊ทธ ์˜ํ–ฅ๋„ ์ปค์„œ, 2016๋…„ ์–€ ๋ฅด์ฟค์€ โ€œ(GAN์€) ์šฐ๋ฆฌ๊ฐ€ ํ•œ๋™์•ˆ ์ƒ๊ฐํ•ด ๋‚ธ ๊ฒƒ๋“ค ์ค‘ ์ตœ๊ณ ์˜ ์•„์ด๋””์–ดโ€๋ผ๊ณ  ๋งํ•œ ์ ๋„ ์žˆ์„ ์ •๋„์˜€์Šต๋‹ˆ๋‹ค.

Image Credit: ์–€ ๋ฅด์ฟค์˜ RI ์„ธ๋ฏธ๋‚˜ โ€˜The Next Frontier in AI: Unsupervised Learningโ€™

VAE (Variational Autoencoders)๋ผ๋“ ๊ฐ€ RBM (Restricted Boltzmann Machines) ๊ฐ™์€ ์ด์ „์˜ ์ƒ์„ฑํ˜• ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ, GAN์€ ๋” ์„ ๋ช…ํ•œ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๋” ๋ณต์žกํ•œ ํŒจํ„ด์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์—ˆ๊ธฐ ๋•Œ๋ฌธ์—, AI ๊ธฐ์ˆ ์˜ ์ƒˆ๋กœ์šด ๊ฐ€๋Šฅ์„ฑ์„ ์—ด์—ˆ๋‹ค๊ณ  ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๋ฌผ๋ก , ๋ฌด์Šจ ๊ธฐ์ˆ ์ด๋“  ํ•œ๊ณ„๊ฐ€ ์žˆ์ฃ  - GAN์ด ๋งž๋‹ฅ๋œจ๋ฆฐ ํ˜„์‹ค์ ์ธ ๋ฌธ์ œ๋Š” ๋ฐ”๋กœ ๋ชจ๋ธ์˜ ๋ถˆ์•ˆ์ •์„ฑ, ๋ชจ๋“œ ๋ถ•๊ดด (Mode Collapse) ๊ฐ™์€ ๊ฒƒ๋“ค์ด์—ˆ์ฃ .

๋””ํ“จ์ „ ๋ชจ๋ธ๋กœ์˜ ์ „ํ™˜

ํฐ ๊ด€์‹ฌ์„ ๋ฐ›์€ GAN์ด์—ˆ์ง€๋งŒ, ์‹œ๊ฐ„์ด ํ˜๋Ÿฌ๊ฐ€๋ฉด์„œ ์ด๋Ÿฐ ๋ถˆ์•ˆ์ •์„ฑ ๋ฌธ์ œ, ๊ทธ๋ฆฌ๊ณ  ๋ชจ๋“œ ๋ถ•๊ดด์˜ ๋ฌธ์ œ ๋“ฑ ํ•™์Šต ์ƒ์˜ ์–ด๋ ค์›€์ด ๋” ์ด์ƒ ๋ฌด์‹œํ•  ์ˆ˜ ์—†๋Š” ๋ฌธ์ œ๊ฐ€ ๋˜์—ˆ์ฃ . ๊ทธ๋Ÿฌ๋˜ ์™€์ค‘์—, 2022๋…„ ์ฆˆ์Œํ•ด์„œ ์ƒˆ๋กœ์šด ๋„์ „์ž์ธ ๋””ํ“จ์ „ ๋ชจ๋ธ (ํ™•์‚ฐ ๋ชจ๋ธ; Diffusion Model)์ด ๋“ฑ์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฑธ ์ ์ง„์ ์ธ ๊ฐœ์„  ๊ณผ์ •์ด๋ผ๋Š” ๊ด€์ ์—์„œ ์ ‘๊ทผํ–ˆ๊ณ , ๊ทธ ๋•Œ๋ฌธ์— ๋” ์•ˆ์ •์ ์ด์—ˆ๊ณ  ํ•™์Šตํ•˜๊ธฐ๋„ ์‰ฌ์› ์Šต๋‹ˆ๋‹ค.

Deep Generative Model ๊ฐ„ ๋น„๊ต. Image Credit: Towards AI

๋””ํ“จ์ „ ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ถœ๋ ฅ๋ฌผ์„ ๊ณ ํ’ˆ์งˆ๋กœ ๋งŒ๋“ค์–ด๋‚ผ ์ˆ˜ ์žˆ์–ด์„œ, ์—ฐ๊ตฌ์ž๋“ค ์ž…์žฅ์—์„œ๋Š” ๊ณจ์นซ๊ฑฐ๋ฆฌ๋Š” ์ ์œผ๋ฉด์„œ ํšจ๊ณผ๊ฐ€ ์ข‹์•„์„œ ๋น ๋ฅด๊ฒŒ ์ฃผ๋ชฉ์„ ๋ฐ›๊ฒŒ ๋˜์—ˆ๊ตฌ์š”. ํ•œ ๋•Œ ์ƒ์„ฑํ˜• ๋ชจ๋ธ์˜ ์Šคํƒ€์˜€๋˜ GAN์€ ๋””ํ“จ์ „ ๋ชจ๋ธ์— ์‚ด์ง (?) ๋ฐ€๋ฆฌ๋Š” ๋“ฏํ•œ ๋Š๋‚Œ์œผ๋กœ ๋Œ€ํ™”์—์„œ ์ ์ฐจ ์‚ฌ๋ผ์ง€๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

The GAN Is Dead; Long Live the GAN!

๊ทธ๋ ‡์ง€๋งŒ, GAN์˜ ๊ฐ€๋Šฅ์„ฑ์„ ์—ฌ์ „ํžˆ ๋ฏฟ๋Š” ์—ฐ๊ตฌ์ž๋“ค์ด ์žˆ์—ˆ์ฃ ! ๋ฐ”๋กœ ๋ฉฐ์น  ์ „์ธ 2025๋…„ 1์›” 9์ผ, โ€˜The GAN Is Dead; Long Live the GAN!โ€™์ด๋ผ๋Š”, ๋Œ€๋‹ดํ•˜๋ฉด์„œ๋„ ์•„์ด๋Ÿฌ๋‹ˆํ•œ ์ œ๋ชฉ์˜ ๋…ผ๋ฌธ์ด ์ง€๊ธˆ GAN์— ๋Œ€ํ•œ ๊ด€์‹ฌ์„ ๋‹ค์‹œ ๋ถˆ๋Ÿฌ์ผ์œผํ‚ค๊ณ  ์žˆ๋Š”๋ฐ์š” - โ€˜GAN์€ ์ฃฝ์—ˆ๋‹ค; GAN ๋งŒ์„ธ!โ€™๋ผ๋Š”, ์–ด์ฐŒ๋ณด๋ฉด ์ด์ƒํ•˜๊ฒŒ ๋“ค๋ฆฌ๋Š” ์ œ๋ชฉ์ด์ฃ ?

Yiwen Huang, Aaron Gokaslan, Volodymyr Kuleshov, ๊ทธ๋ฆฌ๊ณ  James Tompkin์ด ์“ด ์ด ๋…ผ๋ฌธ์€, โ€˜GAN์ด ๋งž๋‹ฅ๋œจ๋ฆฐ ๋ฌธ์ œ์ ๋“ค์ด ๋ณธ์งˆ์ ์ธ ๊ฒฐํ•จ์ด ์•„๋‹ˆ๋ผ ๊ตฌ์‹์˜ ์•„ํ‚คํ…์ฒ˜์™€ ๊ธฐ์ˆ  ๋•Œ๋ฌธ์ด๋‹คโ€™๋ผ๊ณ  ์ฃผ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์—ฐ๊ตฌ์ง„์˜ ์•„์ด๋””์–ด์˜ ํ•ต์‹ฌ์€, โ€˜๋” ์ข‹์€ ์†์‹ค ํ•จ์ˆ˜โ€™์—์š” - ์†์‹ค ํ•จ์ˆ˜๊ฐ€ AI์—์„œ ์–ผ๋งˆ๋‚˜ ์ค‘์š”ํ•œ์ง€๋Š” ๋‹ค๋“ค ์ž˜ ์•„์‹œ์ž–์•„์š”? ๊ฒฐ๊ตญ ๋ณธ์งˆ๋กœ ๋Œ์•„์˜ค๋Š”๊ฐ€๋ด…๋‹ˆ๋‹ค - GAN์ด ์–ผ๋งˆ๋‚˜ ํ•™์Šต์„ ์ž˜ ํ•˜๊ณ  ์žˆ๋Š”์ง€ ์ธก์ •ํ•˜๋Š”, ๋” ๋˜‘๋˜‘ํ•œ ๋ฐฉ๋ฒ•์„ ์ƒˆ๋กœ ๊ณ ์•ˆํ•œ ๊ฑด๋ฐ์š”. ์ด๊ฑธ โ€˜์ƒ๋Œ€์  GAN ์†์‹ค (Relativistic GAN Loss)โ€™๋ผ๊ณ  ๋ถ€๋ฅด๋Š”๋ฐ, GAN์˜ ํ•™์Šต ๊ณผ์ •์„ ๋” ๋ถ€๋“œ๋Ÿฝ๊ฒŒ ๋งŒ๋“ค๊ณ , ๊ฒฐ๊ณผ์ ์œผ๋กœ ์ด์ƒํ•œ ์ถœ๋ ฅ๋ฌผ์ด ๊ฐ‘์ž๊ธฐ ๋‚˜์˜จ๋‹ค๋“ ๊ฐ€, ํ•œ์ •๋œ ์ข…๋ฅ˜์˜ ์ด๋ฏธ์ง€๋งŒ ๋ฐ˜๋ณตํ•ด์„œ ๋‚˜์˜จ๋‹ค๋“ ๊ฐ€ ํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฌธ์ œ๋“ค์ด ๋œ ๋ฐœ์ƒํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

๋˜, ์—ฐ๊ตฌ์ง„์€ GAN ์•„ํ‚คํ…์ฒ˜๋ฅผ โ€˜ํ˜„๋Œ€ํ™”โ€™ํ•˜๊ธฐ๋„ ํ–ˆ๋Š”๋ฐ์š”. ์•„์ฃผ ์‚ฌ์‹ค์ ์ธ ์–ผ๊ตด์„ ์ƒ์„ฑํ•˜๊ธฐ๋กœ ์ž˜ ์•Œ๋ ค์ง„ ๋ชจ๋ธ, StyleGAN2์—์„œ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•ด์„œ, ์ตœ์‹ ์˜ AI ์„ค๊ณ„ ๊ธฐ๋ฒ•๋“ค์„ ๋ฐ˜์˜ํ•ด์„œ ๋” ์ด์ƒ ํ•„์š”ํ•˜์ง€ ์•Š์€ ๋ชจ๋“  ์š”์†Œ๋“ค์„ ๋„๋ ค๋‚ด ๋ฒ„๋ ธ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ResNet์ด๋ผ๋“ ๊ฐ€ โ€˜๊ทธ๋ฃน ํ•ฉ์„ฑ๊ณฑ (Grouped Convolution)โ€™ ๊ฐ™์€ ๋” ๋‚˜์€ ๊ตฌ์„ฑ ์š”์†Œ๋“ค์„ ์ถ”๊ฐ€ํ•ด์„œ R3GAN์ด๋ผ๋Š” ๋” ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ๊ฐ•๋ ฅํ•œ GAN์„ ๋งŒ๋“ค์–ด ๋ƒˆ์Šต๋‹ˆ๋‹ค.

์ด R3GAN์€ ๋” ์ž˜ ์ž‘๋™ํ•  ๋ฟ ์•„๋‹ˆ๋ผ ๋” ๋‹จ์ˆœํ•˜๋‹ค๊ณ  ๋ง์”€๋“œ๋ ธ์ฃ . FFHQ (์‚ฌ๋žŒ ์–ผ๊ตด ๋ฐ์ดํ„ฐ์…‹)๋ผ๋“ ๊ฐ€ CIFAR-10 (์ผ์ƒ์ ์œผ๋กœ ๋ณผ ์ˆ˜ ์žˆ๋Š” ๋ฌผ์ฒด๋“ค์˜ ์ž‘์€ ์ด๋ฏธ์ง€) ๊ฐ™์€ ํ‘œ์ค€ ๋ฒค์น˜๋งˆํฌ์—์„œ, R3GAN์€ ์ผ๋ถ€ ๋””ํ“จ์ „ ๋ชจ๋ธ์„ ํฌํ•จํ•œ ๊ธฐ์กด์˜ ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ๋ชจ์Šต์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ฒŒ๋‹ค๊ฐ€, ํ•™์Šต ์†๋„๋„ ๋” ๋น ๋ฅด๊ณ  ์ปดํ“จํŒ… ํŒŒ์›Œ๋„ ๋œ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.

FFHQ-256 ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ R3GAN์ด ์ƒ์„ฑํ•ด๋‚ธ ์–ผ๊ตด๋“ค. Image Credit: ์˜ค๋ฆฌ์ง€๋„ ๋…ผ๋ฌธ

GAN์ด ๋„ˆ๋ฌด ๊นŒ๋‹ค๋กญ๊ฑฐ๋‚˜ ๊ตฌ์‹์ด๋ผ๊ณ  ์ƒ๊ฐํ•ด์„œ ํ”ผํ•ด์™”๋‹ค๋ฉด, ์ง€๊ธˆ, ๊ทธ๋ฆฌ๊ณ  ์•ž์œผ๋กœ GAN - ์ƒˆ๋กœ์šด GAN์ด๊ฒ ์ฃ  - ์„ ๋‹ค์‹œ ์‹œ๋„ํ•ด ๋ณผ๋งŒํ•œ ์ข‹์€ ์‹œ๊ธฐ๊ฐ€ ์˜ฌ์ง€๋„ ๋ชจ๋ฅด๊ฒ ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋„, GAN์„ ํ™œ์šฉํ•ด์„œ ๋ญ˜ ํ•ด ๋ณผ ์ˆ˜ ์žˆ์„์ง€ ํ•œ ๋ฒˆ ๋‹ค์‹œ ์ƒ๊ฐํ•ด ๋ณผ๋งŒํ•œ ์‹œ๊ฐ„์ด๊ตฌ์š”.

AI ํ˜์‹ , ๊ทธ โ€˜๋ฐ˜๋ณต์ ์ธ (Iterative) ๊ณผ์ •โ€™

์•ž์—์„œ ์‚ดํŽด๋ณธ โ€˜GAN์˜ ๋ถ€ํ™œโ€™. ์—ฌ๊ธฐ์„œ ๋ฐ”๋กœ AI - ๋˜๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ - ํ˜์‹ ์ด ๊ฐ€์ง€๊ณ  ์žˆ๋Š” โ€˜๋ฐ˜๋ณต์ ์ธ (Iterative)โ€™ ํŠน์„ฑ์„ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณผ๊ฑฐ์˜ ์—ฐ๊ตฌ, ํ˜„์žฌ์˜ ์—ฐ๊ตฌ๋“ค์ด ์„œ๋กœ์˜ ์žฅ๋‹จ์ ๋“ค์„ ๋ณต์žกํ•˜๊ฒŒ ์„ž์–ด๊ฐ€๋ฉด์„œ ๋˜ ์ƒˆ๋กœ์šด ์—ฐ๊ตฌ ์ฃผ์ œ๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๊ณ , ๋‹ค์Œ ์„ธ๋Œ€ AI ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ๊ธฐ์ดˆ๋ฅผ ๋งŒ๋“ค์–ด ๋ƒ…๋‹ˆ๋‹ค. ๋ฐ”๋กœ ์ด ๋•Œ๋ฌธ์—, ๋Š์ž„์—†์ด ๋ณ€ํ™”ํ•˜๋Š” AI ์—ฐ๊ตฌ๋ฅผ ์–ด๋–ค ์ˆ˜์ค€์—์„œ๋“  ํŒ”๋กœ์šฐ์—…ํ•˜๋Š” ๊ฒŒ AI ์–ดํ”Œ๋ฆฌ์ผ€์ด์…˜์ด๋‚˜ ์„œ๋น„์Šค๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๋Š” ์ž…์žฅ์—์„œ๋„ ์ค‘์š”ํ•œ ๊ฒƒ์ด ์•„๋‹๊นŒ ํ•ฉ๋‹ˆ๋‹ค.

์ƒˆ๋กœ ๋“ฑ์žฅํ•œ R3GAN์„ ์œ„์‹œํ•ด์„œ, ์ƒˆ๋กœ์šด ์‹œ๋Œ€์˜ ์ƒˆ๋กœ์šด GAN๋“ค์ด ๋˜ ๊ณ„์† ๋“ฑ์žฅํ•˜์ง€ ์•Š์„๊นŒ ํ•˜๋Š”๋ฐ์š”. ๊ณ ํ’ˆ์งˆ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” GAN์˜ ๋Šฅ๋ ฅ, ์ƒ์„ฑํ˜• AI ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ˆ˜์š”๊ฐ€ ํญ์ฆํ•˜๋Š” ์ง€๊ธˆ, ๊ทธ ์–ด๋Š ๋•Œ๋ณด๋‹ค ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ํŠนํžˆ ๊ฐœ์ธ์ •๋ณด์˜ ๋ณดํ˜ธ๋ผ๋“ ๊ฐ€, IP ๋ณดํ˜ธ ๋“ฑ์˜ ๋ฌธ์ œ๋กœ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜๋Š”๊ฒŒ ์–ด๋ ค์šด ์ƒํ™ฉ์—์„œ ๊ทธ ์ค‘์š”์„ฑ์€ ๋”์šฑ ์ปค์ ธ๊ฐˆ ๊ฒƒ์œผ๋กœ ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.

ํŠธ์œ„ํ„ฐ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ (Twitter Library) ๐Ÿฆ

LLM์„ ํฌํ•จํ•ด์„œ, โ€˜์—ฐ๊ตฌโ€™๋ฅผ ์ง€์›ํ•  ๋ชฉ์ ์œผ๋กœ ๋งŒ๋“ค์–ด์ง„ AI ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ํ™œ์šฉํ•œ๋‹ค๋ฉด, ๊ณผํ•™ ์—ฐ๊ตฌ ์ˆ˜ํ–‰์ด ํ›จ์”ฌ ๋” ๊ฐ€์†ํ™”๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜ค๋Š˜์€ ๊ณผํ•™ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด์„œ ๊ณ ์•ˆ๋œ 10๊ฐ€์ง€ AI ์‹œ์Šคํ…œ์„ ์•Œ์•„๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค:

*์•„์ง ํŠœ๋ง ํฌ์ŠคํŠธ ์ฝ”๋ฆฌ์•„ ๊ตฌ๋… ์•ˆ ํ•˜์…จ๋‚˜์š”? ๊ตฌ๋…ํ•ด ์ฃผ์‹œ๋ฉด ๋งค์ฃผ ์ค‘์š”ํ•œ AI ๋‰ด์Šค๋ฅผ ์ •๋ฆฌํ•œ ๋‹ค์ด์ œ์ŠคํŠธ๋ฅผ ๋ฐ›์œผ์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค!

ํŠœ๋ง ํฌ์ŠคํŠธ ํŒ€์ด ์ฝ๊ณ  ์žˆ๋Š” ๊ฒƒ๋“ค ๐Ÿ“

The Focus AI์˜ Will Schenk๊ฐ€ ๊ตฌ๊ธ€์˜ DeepResearch, ์˜คํ”ˆAI์˜ GPT-4o์™€ o1, ๋งˆ์ดํฌ๋กœ์†Œํ”„ํŠธ์˜ Phi-4, Claude, Perplexity, DeepSeek ๋“ฑ ๋‹ค์–‘ํ•œ โ€˜AI ๋ฆฌ์„œ์น˜ ๋„๊ตฌโ€™๋“ค์„ ๋น„๊ตํ•ด ๋ณด๊ณ  ์žˆ๋Š”๋ฐ์š”. โ€˜๋ฐค์ด ์–ด๋‘์šด ์ด์œ ๋Š” ๋ญ˜๊นŒ?โ€™๋ผ๋Š”, ์–ด์ฐŒ๋ณด๋ฉด ๋‹ค์†Œ ์ฒ ํ•™์ ์ด๊ธฐ๋„ ํ•˜๊ณ  ์ƒ๊ฐ๋ณด๋‹ค ๋ณต์žกํ•œ ์งˆ๋ฌธ์„ ๋˜์ง€๊ณ  ๊ทธ ๊ฒฐ๊ณผ๊ฐ€ Will ์ž์‹ ์˜ ์ž…์žฅ์—์„œ ์–ผ๋งˆ๋‚˜ ๋งŒ์กฑ์Šค๋Ÿฌ์šด์ง€ ์ƒ๊ฐํ•ด ๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Will์€ DeepResearch๊ฐ€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค๋ณด๋‹ค ๋” ๊นŠ์ด์žˆ๊ณ , ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๊ณ , ์ฐธ๊ณ  ๋ฌธํ—Œ์ด ์ž˜ ๊ฐ–์ถฐ์ง„ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•œ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ๋ถ„๋„ ํ•œ ๋ฒˆ ์ด๋Ÿฐ ๋น„๊ต ํ•ด ๋ณด์‹œ๋ฉด ์–ด๋–จ๊นŒ์š”?

๊ตฌ๊ธ€, ์• ํ”Œ์„ ๊ฑฐ์นœ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด์ด์ž UX ๋””์ž์ด๋„ˆ์ธ Ben Hylak์ด o1๊ณผ ๊ด€๋ จ๋œ ์ž์‹ ์˜ ๊ฒฝํ—˜์„ ํ’€์–ด๋†“์Šต๋‹ˆ๋‹ค. o1์ด ์ฑ„ํŒ…์„ ์œ„ํ•œ ๋ชจ๋ธ์ด ์•„๋‹ˆ๋ผ ๋ฌธ์ œ ํ•ด๊ฒฐ - ์—ฌ๊ธฐ์„œ๋Š” โ€˜๋ณด๊ณ ์„œ ์ƒ์„ฑโ€™์ด๋ผ๋Š” ํ‘œํ˜„์„ ์ผ๋Š”๋ฐ์š” - ์„ ์œ„ํ•œ ๋ชจ๋ธ๋กœ, o1์˜ ์ž ์žฌ๋ ฅ์„ ์ž˜ ํ™œ์šฉํ•˜๋ ค๋ฉด ์‚ฌ์šฉ์ž ์ž…์žฅ์—์„œ ์–ด๋–ค ์ฃผ์˜ํ•  ์ ์ด๋‚˜ ์Šคํ‚ฌ์„ ๊ฐ–์ถฐ์•ผํ•˜๋Š”์ง€์— ๋Œ€ํ•ด์„œ ๊ธฐ์ˆ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๊ธ€์˜ ๋ง๋ฏธ์—๋Š”, โ€˜์ƒ์„ฑํ˜• AI ๊ธฐ์ˆ ๋กœ ๋ณด๊ณ ์„œ ์ƒ์„ฑ ์„œ๋น„์Šค๋ฅผ ๋งŒ๋“ค ๋•Œ ์ƒ๊ฐํ•ด ๋ณผ ๋งŒํ•œ UI ๊ด€์ ์ด ํŒ(?)์ด๋ผ๊ณ  ํ• ๊นŒ, ๊ทธ๋Ÿฐ ๋‚ด์šฉ๋“ค๋„ ํฌํ•จํ•˜๊ณ  ์žˆ์œผ๋‹ˆ ํ•œ ๋ฒˆ ๊ด€์‹ฌ์žˆ๋Š” ๋ถ„๋“ค ๋ณด์‹œ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹ค.

์นฉ ํ›„์ด์˜Œ์ด ๋ณธ์ธ์˜ ์ €์„œ โ€˜AI Engineeringโ€™์˜ ์—์ด์ „ํŠธ ๊ด€๋ จ ์„น์…˜์„ ๋‹ค์‹œ ํŽธ์ง‘ํ•˜๊ณ  ์—…๋ฐ์ดํŠธํ•ด์„œ ๋‹จ๋… ํฌ์ŠคํŠธ โ€˜Agentsโ€™๋ฅผ ๋งŒ๋“ค์—ˆ๋„ค์š”. ์นฉ ํ›„์ด์˜Œ ๋ฒ„์ „์˜ โ€˜Agent์˜ ๋ชจ๋“  ๊ฒƒโ€™ ์ •๋„๋กœ ์ƒ๊ฐํ•˜๊ณ  ์ผ๋…ํ•ด ๋ณผ๋งŒํ•œ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค.

์•จ๋ŸฐAI ์—ฐ๊ตฌ์†Œ์˜ Nathan Lambert๊ฐ€ NeurIPS์—์„œ ํ–ˆ๋˜ LLM ๊ด€๋ จ ํŠœํ† ๋ฆฌ์–ผ์„ ๋‹ค์‹œ ๋…นํ™”ํ•œ ๋น„๋””์˜ค, ๊ทธ๋ฆฌ๊ณ  ์ •๋ฆฌํ•œ ๊ธ€์ธ๋ฐ์š”. ์ž‘๋…„ ํ›„๋ฐ˜๋ถ€ํ„ฐ ํฐ ์ด์Šˆ๊ฐ€ ๋˜๊ณ  ์žˆ๋Š” Post-Training ๊ด€๋ จํ•ด์„œ ๋งŽ์€ ๋‚ด์šฉ์ด ๋‹ด๊ฒจ ์žˆ์Šต๋‹ˆ๋‹ค.

์ƒˆ๋กœ ๋‚˜์˜จ, ์ฃผ๋ชฉํ•  ๋งŒํ•œ ์—ฐ๊ตฌ ๋…ผ๋ฌธ

๊ธˆ์ฃผ์˜ Top Pick!

Top Pick์— ํ•ด๋‹นํ•˜๋Š” ๋…ผ๋ฌธ๋“ค์ด ์ƒ๋‹นํžˆ ๋งŽ๋„ค์š” ^.^;

  • Sky-T1: Train Your Own O1 Preview Model Within $450
    ์ด ๋…ผ๋ฌธ์€ ์ถ”๋ก ๊ณผ ์ฝ”๋”ฉ ์ž‘์—…์„ ๋ชฉ์ ์œผ๋กœ 32B ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๊ณผ์ •์„ ํ†ตํ•ด์„œ โ€˜๊ณ ์„ฑ๋Šฅ ์ถ”๋ก  ๋ชจ๋ธ์˜ ๊ฒฝ์ œ์„ฑโ€™์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • RStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking

    ๋ชฌํ…Œ์นด๋ฅผ๋กœ ํŠธ๋ฆฌ ์„œ์น˜์™€ ๋ฐ˜๋ณต์ ์ธ ์ž๊ธฐ ๊ฐœ์„  (Self-Improvement) ๊ธฐ๋ฒ•์„ ํ†ตํ•ด์„œ ์†Œํ˜• ๋ชจ๋ธ๋“ค์ด ์ˆ˜ํ•™์  ์ถ”๋ก ์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฑธ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • Test-time Computing: From System-1 Thinking to System-2 Thinking

    ๊ฒฌ๊ณ ํ•˜๊ฒŒ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ, ์ง๊ด€์ ์ธ ์ „๋žต, ๊ทธ๋ฆฌ๊ณ  ์‹ฌ์‚ฌ์ˆ™๊ณ ํ•˜๋Š” ์ „๋žต์„ ๊ฒฐํ•ฉํ•ด์„œ AI์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • Towards System 2 Reasoning in LLMs: Learning How to Think with Meta Chain-of-Thoughts

    ๋ณต์žกํ•œ ๋ฌธ์ œ ํ•ด๊ฒฐ ์ž‘์—…์—์„œ์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ, ๋ฐ˜๋ณต์  ํƒ์ƒ‰๊ณผ ๊ฒ€์ฆ์„ ํ•  ์ˆ˜ ์žˆ๋Š” Meta-CoT ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • REINFORCE++: A Simple and Efficient Approach for Aligning Large Language Models
    PPO์—์„œ ์˜๊ฐ์„ ๋ฐ›์€ ๊ธฐ์ˆ ๋“ค์„ โ€˜REINFORCEโ€™ ํ”„๋ ˆ์ž„์›์— ํ†ตํ•ฉ, RLHF๋ฅผ ๊ฐœ์„ ํ–ˆ๊ณ , ์ด๋Ÿฐ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด์„œ ๋น„ํ‰๊ฐ€ ๋„คํŠธ์›Œํฌ(Critic Network) ์—†์ด๋„ ๋” ๋น ๋ฅด๊ณ , ์•ˆ์ •์ ์ด๋ฉฐ, ํšจ์œจ์ ์ธ Alignment๋ฅผ ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฑธ ๋ณด์—ฌ์คฌ์Šต๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • Cosmos World Foundation Model Platform for Physical AI
    ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ๋Œ€์ƒ์œผ๋กœ, ๋ฌผ๋ฆฌ ๋ฒ•์น™์„ ๊ณ ๋ คํ•œ ๋Œ€๊ทœ๋ชจ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ํ†ตํ•ด์„œ ๋กœ๋ณดํ‹ฑ์Šค ์‹œ์Šคํ…œ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ชจ๋ธ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

์ถ”๋ก  ๋ฐ ์ˆ˜ํ•™์  ์—ญ๋Ÿ‰

  • Search-o1: Agentic Search-Enhanced Large Reasoning Models
    ์ถ”๋ก  ๋ชจ๋ธ์„ ์œ„ํ•œ RAG์„ ๋„์ž…, ์™ธ๋ถ€ ์ง€์‹์„ ํ†ตํ•ฉํ•ด์„œ ๋ณต์žกํ•œ ๋„๋ฉ”์ธ์—์„œ ๋‹ต๋ณ€์˜ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • BoostStep: Boosting Mathematical Capability of Large Language Models via Improved Single-Step Reasoning
    ์ˆ˜ํ•™์  ๊ณผ์ œ๋ฅผ ์œ„ํ•œ ๋‹จ๊ณ„๋ณ„ ์ถ”๋ก ์„ ๊ฐœ์„ ํ•ด์„œ, ์œ ์‚ฌ์„ฑ์ด ๋‚ฎ๊ณ  ์–ด๋ ค์šด ๋ฒค์น˜๋งˆํฌ์—์„œ ์ •ํ™•๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • URSA: Understanding and Verifying Chain-of-thought Reasoning in Multimodal Mathematics
    ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ž‘์—…์„ ์œ„ํ•œ CoT ์ถ”๋ก ์— ์ค‘์ ์„ ๋‘๊ณ , ์ˆ˜ํ•™์ ์ธ ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ ๊ฒฌ๊ณ ํ•œ ํ”„๋ ˆ์ž„์›์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • DOLPHIN: Closed-Loop Open-Ended Auto-Research through Thinking, Practice, and Feedback
    ์•„์ด๋””์–ด ์ƒ์„ฑ, ๊ฒ€์ฆ, ๊ฐœ์„ ์„ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ˜๋ณต์  ํ”ผ๋“œ๋ฐฑ ๋ฃจํ”„๋ฅผ ํ†ตํ•ด์„œ ์ž๋™์ ์ธ ์—ฐ๊ตฌ ์ˆ˜ํ–‰ ๊ณผ์ •์„ ํ˜์‹ ํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

Robotics and Physical AI

  • OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives
    ๋กœ๋ด‡ ์กฐ์ž‘์„ ์œ„ํ•œ ๋น„์ „-์–ธ์–ด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜๋ฉด์„œ, ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ โ€˜Zero-shot ์ผ๋ฐ˜ํ™” (Generalization)โ€™๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

RAG (Retrieval-Augmented Generation)

  • VideoRAG: Retrieval-Augmented Generation over Video Corpus
    ๋น„๋””์˜ค ๊ธฐ๋ฐ˜์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์‘๋‹ต ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด์„œ ์‹œ๊ฐ์  ๊ฒ€์ƒ‰๊ณผ ํ…์ŠคํŠธ ๊ฒ€์ƒ‰์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

  • Personalized Graph-Based Retrieval for Large Language Models enriches
    ๊ฐœ์ธํ™”๋œ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด์„œ, ์‚ฌ์šฉ์ž ์ค‘์‹ฌ์˜ ์ง€์‹ ๊ทธ๋ž˜ํ”„๋ฅผ ํ†ตํ•ฉ, ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
    โ€”> [๋” ๋ณด๊ธฐ]

  • GeAR: Generation Augmented Retrieval
    ์„ธ๋ฐ€ํ•œ ํ…์ŠคํŠธ ๋‹จ์œ„๋ฅผ ์ฐพ๊ณ  ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด์„œ ์ด์ค‘ ์ธ์ฝ”๋”(Bi-Encoder) ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๊ฒ€์ƒ‰๊ณผ ์ƒ์„ฑ์„ ์—ฐ๊ฒฐํ•ฉ๋‹ˆ๋‹ค. โ€”> [๋” ๋ณด๊ธฐ]

*๋ฆฌ๋ทฐ๋ฅผ ๋‚จ๊ธฐ์‹œ๋ ค๋ฉด ๋กœ๊ทธ์ธํ•˜์‹œ๊ฑฐ๋‚˜ ๊ตฌ๋…ํ•ด ์ฃผ์„ธ์š”. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค!

์ฝ์–ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํ”„๋ฆฌ๋ฏธ์—„ ๊ตฌ๋…์ž๊ฐ€ ๋˜์–ด์ฃผ์‹œ๋ฉด ํŠœ๋ง ํฌ์ŠคํŠธ ์ฝ”๋ฆฌ์•„์˜ ์ œ์ž‘์— ํฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค!

Reply

or to participate.