[Review] Meta-Learning: Learning to Learn Fast

YSLยท2023๋…„ 8์›” 4์ผ
0

Review

๋ชฉ๋ก ๋ณด๊ธฐ
6/7
post-thumbnail

๐Ÿ“ Meta-Learning: Learning to Learn Fast

โ—๏ธ๊ฐœ๋…์„ ์ •๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์„ฑํ•œ ๊ธ€๋กœ, ๋‚ด์šฉ์ƒ ์ž˜๋ชป๋œ ๋ถ€๋ถ„์ด ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์  ์ฐธ๊ณ  ๋ฐ”๋ž๋‹ˆ๋‹ค.

์ด๋ฒˆ์ฃผ๋Š” Meta-Learning์— ๋Œ€ํ•ด ์ •๋ฆฌ๋œ ๋ธ”๋กœ๊ทธ๋ฅผ ์ฝ๊ณ  ์ „๋ฐ˜์ ์ธ ๋‚ด์šฉ์„ ๊ณต๋ถ€ํ–ˆ๋‹ค. ์›Œ๋‚™ ๋‚ด์šฉ์ด ๋ฐฉ๋Œ€ํ•ด ์ƒˆ๋กœ ๊ธ€์„ ์ž‘์„ฑํ•˜๋ฉด ์‹œ๊ฐ„์ด ์˜ค๋ž˜ ๊ฑธ๋ฆด ๊ฒƒ ๊ฐ™์•„ ๋‚ด์šฉ ์ •๋ฆฌ๋ฅผ ์œ„ํ•ด ๋งŒ๋“ค์—ˆ๋˜ PPT ์Šฌ๋ผ์ด๋“œ๋ฅผ ์ฒจ๋ถ€ํ•  ๊ฒƒ์ด๋‹ค.


  • FOMAML

    L(1)L^{(1)}, L(0)L^{(0)} : ๊ฐ™์€ task ๋‚ด์—์„œ ๋‹ค๋ฅธ batch๋กœ ๊ตฌํ•œ ์†์‹ค

    • ๊ธฐ์กด : (tโˆ’1)(t-1)๋ฒˆ์งธ๊นŒ์ง€ loss๊ฐ€ ์—…๋ฐ์ดํŠธ๋˜๋Š” ๊ณผ์ •์„ ์—ฐ์†์ ์œผ๋กœ ๊ตฌํ•˜๊ณ  tt๋ฒˆ์งธ task์— ๋Œ€ํ•ด ์ƒˆ๋กœ์šด ๋ฐฐ์น˜ ์ƒ˜ํ”Œ์„ ์‚ฌ์šฉํ•ด ํ•ด๋‹น task์— ๋Œ€ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ตฌํ•จ
    • FOMAML : (tโˆ’1)(t-1)๋ฒˆ์งธ task์—์„œ ๊ตฌํ•œ loss๋งŒ ์‚ฌ์šฉํ•ด tt๋ฒˆ์งธ task์— ๋Œ€ํ•ด ์ƒˆ๋กœ์šด ๋ฐฐ์น˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•ด ํ•ด๋‹น task์— ๋Œ€ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ตฌํ•จ
  • Reptile
    ๊ฐ task์— n๋ฒˆ ์ •๋„์˜ SGD ๋˜๋Š” ๋ฏธ๋‹ˆ๋ฐฐ์น˜ SGD๋ฅผ ์‚ฌ์šฉํ•ด ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๊ณ , ์ด๋ ‡๊ฒŒ ์–ป์€ ๊ฐ task์˜ ๐œƒโ€™์™€ ์ดˆ๊ธฐ ๐œƒ ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ค„์ด๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๐œƒ ์ตœ์ ํ™”ํ•จ
    โ‡” Task๊ฐ€ ๊ฐ๊ฐ์˜ ๋งค๋‹ˆํด๋“œ ๊ณต๊ฐ„ ์ƒ์— ์žˆ๋‹ค๊ณ  ๊ฐ€์ •ํ–ˆ์„ ๋•Œ, ๊ฐ ๋งค๋‹ˆํด๋“œ ๊ณต๊ฐ„๊ณผ์˜ ๊ฑฐ๋ฆฌ๊ฐ€ ํ‰๊ท ์ ์œผ๋กœ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์œ„์น˜์— ์ดˆ๊ธฐ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์„ธํŒ…๋˜๋„๋ก ํ•™์Šต


References

๐Ÿ“ [DL] Meta-Learning: Learning to Learn Fast
ใ„ด ์˜๋ฌธ ๋ธ”๋กœ๊ทธ๋ฅผ ๊ทธ๋Œ€๋กœ ํ•ด์„ํ•ด์ฃผ์…”์„œ ๊ณต๋ถ€ํ•  ๋•Œ ํฐ ๋„์›€์ด ๋˜์—ˆ๋‹ค !
๐Ÿ“ Few-shot Learning
๐Ÿ“ [Deep Learning] Few shot Learning, Meta learning ๊ฐœ๋… ์ด์ •๋ฆฌ
๐Ÿ“ Meta-Learning
๐Ÿ“ [10์ฃผ์ฐจ] (MAML) Model-agnostic Meta Learning for Fast Adaptation of Deep Networks ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ
๐Ÿ“ Meta Learning With Medical Imaging and Health Informatics Applicaitons(CH5. Model-based meta learning)
๐Ÿ“ Meta-Learning: An Overview and Applications
๐Ÿ“ Meta-Learning๊ณผ MAML์˜ ๊ฐœ๋… ์ •๋ฆฌ
๐Ÿ“ From zero to research - An introduction to Meta-learning
๐Ÿ“ [Meta-Learning] 1. ๋ฉ”ํƒ€, ๋ฉ”ํƒ€๋Ÿฌ๋‹์ด๋ž€ ๋ญ˜๊นŒ?
๐Ÿ“ [๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Matching Networks for one shot learning
๐Ÿ“ [๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Learning to Compare: Relation Network for Few-Shot Learning
๐Ÿ“ Meta-003, Meta-Learning with Memory-Augmented Neural Networks (2016-JMLR)
๐Ÿ“ Long Short-Term Memory (LSTM) ์ดํ•ดํ•˜๊ธฐ
๐Ÿ“ Neural Turing Machines ๋ถ„์„
๐Ÿ“ [meta] (paper 1) Meta learning with Memory Augmented Neural Networks
๐Ÿ“ Meta-Learning with Memory Augmented Neural Networks
๐Ÿ“ Meta-Learning 2. Optimization as a model for few-shot learning
๐Ÿ“ On First-Order Meta-Learning Algorithms
๐Ÿ“ [meta] Meta Learning ์†Œ๊ฐœ
๐Ÿ“ META_LEARNING IS ALL YOU NEED
๐Ÿ“ Meta learning with memory augmented neural network


Meta-Learning์— ๋Œ€ํ•œ ๊ฐœ๋…์ด ์•„์ง ๋ชจํ˜ธํ•˜๊ธฐ๋„ ํ•˜๊ณ , ํŠนํžˆ Model-based approach์˜ MANN์—์„œ ์ดํ•ด๊ฐ€ ์•ˆ๋˜๋Š” ๋‚ด์šฉ์ด ๋งŽ์•„ CS294๋ฅผ ์ถ”๊ฐ€์ ์œผ๋กœ ๊ณต๋ถ€ํ•  ์˜ˆ์ •์ด๋‹ค.

1๊ฐœ์˜ ๋Œ“๊ธ€

comment-user-thumbnail
2023๋…„ 8์›” 6์ผ

์—ด์‹ฌํžˆ ํ•˜์‹œ๋„ค์š”^^ ์ข‹์€๊ธ€ ์ž˜๋ณด๊ณ  ๊ฐ‘๋‹ˆ๋‹ค~

๋‹ต๊ธ€ ๋‹ฌ๊ธฐ