2024. 10. 8. 00:57ใArtificialIntelligence/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

๊ทธ๋๋ ์๋ ์ ๋ ผ๋ฌธ์ ์จ๋ดค๋ ๋ถ์ผ๋ผ, ์ฌ๋ฏธ์๊ฒ ๋ค์์ต๋๋ค. :)

Dataset Distillation Workshop Introduction


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 Julyโ23, Congrats to Arushi for ICCV and Wei-Hong for IJCV acceptances. Septโ22, 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
๋งค์ฐ ํฐ ๊ท๋ชจ์ ์ด๋ฏธ์ง์ ์ ์๊ฒ condense ํ๋ ๊ฒ
Generative model - systhetic dataset - useful for training
SOTA ๋ชจ๋ธ๋ค- compution ์๊ตฌ๊ฐ ๋๊ณ ์๋ค (Heavy Deep Networks)
Training models -> OpenAI - 5 gigawatt datacenter (์๋์ง๋ ๋ง์ด ๋ ๋ค)
Data efficient training
Gradient update
DD - ๋งค์ฐ ํฐ ๋ฐ์ดํฐ ์ ์ผ๋ก๋ถํฐ ์์ ๋ฐ์ดํฐ์ ์ ํฉ์ฑํ๋ ๊ฒ
By loss function
ํญ์ ๋ฐ์ดํฐ์ ์ distillationํ๋๊ฒ ์ฌ๋ฐ๋ฅธ ๊ฒ์ ์๋!
- ์ธ์ DD๊ฐ ์ ์ฉํ๊ฐ?
- Limited memory ์ํฉ์์ DD๊ฐ ์ ์๋ฏธ
- Coreset์ downstream task์ ๋ง๊ฒ ์ป์ด์ง๋๊ฒ ์๋
- Knowledge Distillation
- Teacher network -> student network
- ๋ฐ์ดํฐ ์ ์ ํฉ์ฑํ๋ ๊ฒ์ ์๋๋ค. (๋๊ฒจ์ฃผ๋ ๊ฒ ๋ฟ)
- ์์ฑ ๋ชจ๋ธ -> ๋ณ๊ฐ์ ์ํ์ ์์ฑํ๋ ๊ฒ (์ฌ์ค์ ์ธ ์ด๋ฏธ์ง)
- ๋ฐ์ดํฐ์
๋์คํธ๋ ์ด์
์ learning images (imformatic for training)
- Matching DD
- Similar path๋ฅผ ๊ฐ๋๋ก ํ๋ ๊ฒ
- Small synthetic set
- Compute error -> Loss func + gradient descent / update model - Back-propagate -> training dataset
- Inner loop optimization์ ๋งค์ฐ ๋น ๋ฅด๊ฒ ํ๋ ์ต๊ทผ ๊ฒฝํฅ
- Parameter matching DD
- Similar parameters -> similar performance ์์ํ๋ ๊ฒ
- ๋ด๊ฐ ํ ๊ฒ! + ๋ฌธ์ ๊ฐ ์กด์ฌ (difficult to optimize)
- curriculum ๋ฐฉ์
- Each training step T
- Still expensive
- Distribution Matching in DD / ์ด๊ฒ๋ ์ฝ์๋ ๋ ผ๋ฌธ
- ๊ทธ๋ํ๋ ์ง์ง ๋น์ทํ๋ค . . ๋ค์ ์ฝ์ด๋ด์ผ๊ฒ ๋ค.
- DD with GAN
- Label distillation - ์ฌ๋ฏธ์์ด ๋ณด์ธ๋ค.!
- How to Parameter distillation?
- Informative images
- Distillation performance ๋น๊ตํ ๋ํ
- IPC๊ฐ ์ฆ๊ฐํ ์๋ก ์๋ ดํ๋ ์ฑ๋ฅ
- ์์ง ์ฐ๋ฆฌ๊ฐ ํด๊ฒฐํด์ผ ํ ๋ฌธ์ ๊ฐ ๋จ์์๋ค + another questions
- Beyond image classification -> ํ์ฌ๋ ๋ถ๋ฅ์๋ง ์ง์คํ๋ ์ค
- Multi-modal models (vision-language task) ์์ง ๊ฐ์ผํ ๊ธธ์ด ๋ฉ๋ค.
- Images and some captions with that
'ArtificialIntelligence > ECCV2024' ์นดํ ๊ณ ๋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 |