ECCV 2024 DAY2 - Dataset Distillation Workshop (1)

2024. 10. 8. 00:57ใ†ArtificialIntelligence/ECCV2024

 

 

 

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

 

https://dd-challenge-main.vercel.app/#/

 

Dataset Distillation Challenge

 

dd-challenge-main.vercel.app

 

 

 

 

๊ทธ๋ž˜๋„ ์ž‘๋…„์— ๋…ผ๋ฌธ์„ ์จ๋ดค๋˜ ๋ถ„์•ผ๋ผ, ์žฌ๋ฏธ์žˆ๊ฒŒ ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. :) 

 

 

 

๋ชฉ์ฐจ

 

 

 

Dataset Distillation Workshop Introduction 

 

๋ชฉ์ฐจ

 

 

 

1st An Introduction to Dataset Distillation

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

1st An Introduction to dataset distillation - Hakan Bilen 

https://homepages.inf.ed.ac.uk/hbilen/

 

Hakan Bilen @ ed.ac.uk

Dr. Hakan Bilen Google Scholar GitHub School of Informatics University of Edinburgh 10 Crichton Street Edinburgh EH8 9AB NEWS / ACTIVITY Julyโ€™23, Congrats to Arushi for ICCV and Wei-Hong for IJCV acceptances. Septโ€™22, NeurIPS paper accepted, congrats t

homepages.inf.ed.ac.uk

 

https://github.com/VICO-UoE/DatasetCondensation

 

GitHub - VICO-UoE/DatasetCondensation: Dataset Condensation (ICLR21 and ICML21)

Dataset Condensation (ICLR21 and ICML21). Contribute to VICO-UoE/DatasetCondensation development by creating an account on GitHub.

github.com

 

 

 

* Summary

๋งค์šฐ ํฐ ๊ทœ๋ชจ์˜ ์ด๋ฏธ์ง€์…‹์„ ์ž‘๊ฒŒ condense ํ•˜๋Š” ๊ฒƒ 

Generative model - systhetic dataset - useful for training 

SOTA ๋ชจ๋ธ๋“ค- compution ์š”๊ตฌ๊ฐ€ ๋Š˜๊ณ  ์žˆ๋‹ค (Heavy Deep Networks)

Training models -> OpenAI - 5 gigawatt datacenter (์—๋„ˆ์ง€๋„ ๋งŽ์ด ๋“ ๋‹ค) 

 

Data efficient training 

Gradient update 

DD - ๋งค์šฐ ํฐ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ๋ถ€ํ„ฐ ์ž‘์€ ๋ฐ์ดํ„ฐ์…‹์„ ํ•ฉ์„ฑํ•˜๋Š” ๊ฒƒ 

By loss function 

 

ํ•ญ์ƒ ๋ฐ์ดํ„ฐ์…‹์„ distillationํ•˜๋Š”๊ฒŒ ์˜ฌ๋ฐ”๋ฅธ ๊ฒƒ์€ ์•„๋‹˜!

  • ์–ธ์ œ DD๊ฐ€ ์œ ์šฉํ•œ๊ฐ€? 
  • Limited memory ์ƒํ™ฉ์—์„œ DD๊ฐ€ ์œ ์˜๋ฏธ
  • Coreset์€ downstream task์— ๋งž๊ฒŒ ์–ป์–ด์ง€๋Š”๊ฒŒ ์•„๋‹˜
  • Knowledge Distillation
  • Teacher network -> student network 
  • ๋ฐ์ดํ„ฐ ์…‹์„ ํ•ฉ์„ฑํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค. (๋„˜๊ฒจ์ฃผ๋Š” ๊ฒƒ ๋ฟ) 
  • ์ƒ์„ฑ ๋ชจ๋ธ -> ๋ณ„๊ฐœ์˜ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ (์‚ฌ์‹ค์ ์ธ ์ด๋ฏธ์ง€)
  • ๋ฐ์ดํ„ฐ์…‹ ๋””์Šคํ‹ธ๋ ˆ์ด์…˜์€ learning images (imformatic for training)
     
  • Matching DD
  • Similar path๋ฅผ ๊ฐ–๋„๋ก ํ•˜๋Š” ๊ฒƒ
  • Small synthetic set
  • Compute error -> Loss func + gradient descent / update model - Back-propagate -> training dataset
     
  • Inner loop optimization์„ ๋งค์šฐ ๋น ๋ฅด๊ฒŒ ํ•˜๋Š” ์ตœ๊ทผ ๊ฒฝํ–ฅ
  • Parameter matching DD 
  • Similar parameters -> similar performance ์˜ˆ์ƒํ•˜๋Š” ๊ฒƒ 
  • ๋‚ด๊ฐ€ ํ•œ ๊ฒƒ! + ๋ฌธ์ œ๊ฐ€ ์กด์žฌ (difficult to optimize)

  • curriculum ๋ฐฉ์‹ 
  • Each training step T
  • Still expensive
     
  • Distribution Matching in DD / ์ด๊ฒƒ๋„ ์ฝ์—ˆ๋˜ ๋…ผ๋ฌธ 
  • ๊ทธ๋ž˜ํ”„๋„ ์ง„์งœ ๋น„์Šทํ•˜๋„ค . . ๋‹ค์‹œ ์ฝ์–ด๋ด์•ผ๊ฒ ๋‹ค.

  • DD with GAN 
  • Label distillation - ์žฌ๋ฏธ์žˆ์–ด ๋ณด์ธ๋‹ค.!
  • How to Parameter distillation? 
  • Informative images 
  • Distillation performance ๋น„๊ตํ•œ ๋„ํ‘œ
     
  • IPC๊ฐ€ ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ์ˆ˜๋ ดํ•˜๋Š” ์„ฑ๋Šฅ 
  • ์•„์ง ์šฐ๋ฆฌ๊ฐ€ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ฌธ์ œ๊ฐ€ ๋‚จ์•„์žˆ๋‹ค + another questions 
  • Beyond image classification -> ํ˜„์žฌ๋Š” ๋ถ„๋ฅ˜์—๋งŒ ์ง‘์ค‘ํ•˜๋Š” ์ค‘ 
  • Multi-modal models (vision-language task) ์•„์ง ๊ฐ€์•ผํ•  ๊ธธ์ด ๋ฉ€๋‹ค.
  • Images and some captions with that

 

 

 

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

ECCV 2024 DAY3 - Demo Session  (7) 2024.10.18
ECCV 2024 DAY2 - Dataset Distillation Workshop (2)  (2) 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
ECCV 2024 DAY1 - Registration  (0) 2024.09.30