ECCV 2024 DAY2 - Dataset Distillation Workshop (2)

2024. 10. 8. 20:30ใ†ArtificialIntelligence/ECCV2024

 

 

 

ECCV 2024 Day2 
Sometimes Less is More: The First Dataset Distillation Challenge

 

Flavors of Distillation 

ํ—ฌ๋ฆ„ํ™€์ธ  ยท ๋ฎŒํ—จ ๊ณต๋Œ€

 

distillation๊ณผ ์—ฐ๊ด€๋œ ์ฃผ์ œ๋“ค์„ ์ฐจ๋ก€๋กœ ์—ฐ๊ตฌ์™€ ํ•จ๊ป˜ ์†Œ๊ฐœํ•ด์ฃผ์…”์„œ, ์ดํ•ด๊ฐ€ ์ž˜ ๋˜์—ˆ๋‹ค.

 

 

 

 

 

 

 

 

dream

 

 

 

 

 

 

 

 

 

 

์•„๋ž˜์™€ ๊ฐ™์ด Part ๋ณ„๋กœ ์—ฐ๊ตฌํ•˜์‹  ๋‚ด์šฉ ยท ๋…ผ๋ฌธ์„ ํ•จ๊ป˜ ์„ค๋ช…ํ•ด์ฃผ์…จ๋‹ค.

Next step์œผ๋กœ ๋„˜์–ด๊ฐˆ ๋•Œ, ์™œ ํ•ด๋‹น ๋ฐฉ๋ฒ•๋ก ์ด ๋„์ž…๋˜์—ˆ๋Š”์ง€ ์ œ์‹œํ•ด์ฃผ์…”์„œ,

์ดํ•ดํ•˜๋ฉด์„œ ๋“ฃ๊ธฐ ์ข‹์•˜๋‹ค.! :) 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

* Summary

(1) Data distillation - (2) test data curation - (3) knowledge distillation - (4) General knowledge distillation

 

์ž‘์€ training set -> small samples synth 

  • Data dream 
  • Fine tuning the generator
  • LoRA Adapter
  • Different adapter for each class of training set
  • Generates high quality images + SOTA Results 

 

Data dream summary 

Few shot guided dataset condensation 

 

(2) Test data curation - EgoCVR

Composed Video Retrieval

  • ํ˜„์žฌ์˜ ๋ฌธ์ œ + new method 
  • Input - video -> text instruction + captioner -> text to video retrieval
  • Summary 

 

(3) Knowledge Distillation - teacher and student models 

Large model + compressing model 

Via feature distillation 

 

 (4) General knowledge distillation

์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต: Yes

์ฒ˜์Œ์—๋Š” KD๋กœ ์‹œ์ž‘ -> ์„ฑ๋Šฅ์ด ๋‚ฎ์€ ๋ชจ๋ธ์ด๋”๋ผ๋„์ƒˆ๋กœ์šด ๋ถ€๋ถ„์„ ํ•™์ƒ ๋ชจ๋ธ์ด ์กฐ๊ธˆ ์–ป์–ด๊ฐ 

 

๋…ผ๋ฌธ๊ณผ ์—ฐ๊ตฌ ๋ฐฉ์‹์˜ ๋™๊ธฐ + method + ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด ๊ฐ„๋žตํ•˜๊ฒŒ flow๋ฅผ ์„ค๋ช…ํ•ด์ฃผ์…”์„œ 

์—ฐ๊ตฌ์˜ ํฐ ๊ทธ๋ฆผ์„ ํŒŒ์•…ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์ˆ˜์›”ํ–ˆ๋‹ค.

Summary๋กœ ๊ฐ„๋žตํ•˜๊ฒŒ ์†Œ๊ฐœํ•˜๋Š” ๊ฒƒ (gains๋ฅผ) 

 

 

 

 

  • Panel discussion 

์ด๋ฏธ์ง€์—๋งŒ ์ง‘์ค‘ํ•œ ํ˜„์žฌ - just on the image based

DD์˜ ๋ฏธ๋ž˜ -> ๋‹ค๋ฅธ modality์— ์–ด๋–ป๊ฒŒ extend ํ•  ๊ฒƒ์ธ๊ฐ€? 

Or ์‹ค์„ธ๊ณ„์˜ ๋ฐ์ดํ„ฐ์— ์ ์šฉํ•  ๊ฒƒ์ธ๊ฐ€? ํƒœ๋ถˆ๋Ÿฌ ๋ฐ์ดํ„ฐ์…‹ / ๋น„๋””์˜ค - very new topic / 

๋งŽ์€ challenges ๊ฐ€ ๋‚จ์•„์žˆ๋‹ค - cross architecture performance 

 

+ Multi-labeled - More complex data 

  • Poster session - panel discussions - future of dataset distillation 
  • Focused on the next step (multi-modality or more complex datasets)

 

 

 

โ˜•๏ธCoffee break ยท poster session 

์ปคํ”ผ ํƒ€์ž„ -

 

 

 

 

 

 

๐Ÿค”

์ž‘๋…„์— ๋…ผ๋ฌธ ์“ฐ๋ฉด์„œ, ๋ชจ๋ธ ๋Œ๋ฆด ๋•Œ ๊ถ๊ธˆํ–ˆ๋˜ ์ ์ด ์žˆ์—ˆ๋Š”๋ฐ 

ํฌ์Šคํ„ฐ ๊ตฌ๊ฒฝํ•˜๋‹ค๊ฐ€ ๋˜‘๊ฐ™์€ ํ˜„์ƒ์ด ๋ณด์ด๊ธธ๋ž˜, ์ €์ž๋ถ„๊ป˜ ์งˆ๋ฌธํ•ด๋ดค๋‹ค.!

Condensation์„ ์‹œํ‚ค๋ฉด ์ €๋ ‡๊ฒŒ ์ด๋ฏธ์ง€์— ๊ฒฉ์ž๋ฌด๋Šฌ ํŒจํ„ด์ด ์ƒ๊ธฐ๋Š”๋ฐ, ์ฒ˜์Œ์—๋Š” ์ด๊ฒŒ ์˜ค๋ฅ˜์ธ ์ค„ ์•Œ์•˜๊ฑฐ๋“ฑ์š”

์—ฌ์ญค๋ณด๋‹ˆ Convolution ์—ฐ์‚ฐ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋„คํŠธ์›Œํฌ์—์„œ, ํ•„ํ„ฐ ๋•Œ๋ฌธ์— ๋“œ๋Ÿฌ๋‚˜๋Š” ํ˜„์ƒ ๊ฐ™๋‹ค๊ณ  ์•Œ๋ ค์ฃผ์…จ๋‹ค. 

+ ViT์ฒ˜๋Ÿผ ๋‹ค๋ฅธ ๋„คํŠธ์›Œํฌ์—์„œ๋Š” ์ €๋Ÿฐ ํ˜•ํƒœ๊ฐ€ ๋“œ๋Ÿฌ๋‚˜์ง€ ์•Š๋Š”๋‹ค๊ณ  ํ•˜์‹œ๋”๋ผ๊ตฌ์š”,, 

์›์ธ์„ ์ด์ œ์„œ์•ผ ์•Œ๊ฒŒ ๋˜๋‹ค๋‹ˆ,, ๊ฝค๋‚˜ ์‹ ๊ธฐํ•˜๊ณ  ์žฌ๋ฏธ์žˆ์—ˆ๋‹ค. :) 

 

 

 

 

 

'ArtificialIntelligence > ECCV2024' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

ECCV 2024 Day3 - Oral: Recognition  (0) 2024.11.13
ECCV 2024 DAY3 - Demo Session  (7) 2024.10.18
ECCV 2024 DAY2 - Dataset Distillation Workshop (1)  (0) 2024.10.08
ECCV 2024 DAY1 - Quantum Computer Vision (2)  (2) 2024.10.08
ECCV 2024 DAY1 - Quantum Computer Vision (1)  (0) 2024.10.08