06-24 Live Session

2 minute read

CS & Calculus background

Domain Generalization

  • The idea of Domain Generalization is to learn from one or multiple training domains, to extract a domain-agnostic model which can be applied to an unseen domain.

Self-Supervised-Learning

  • Self-Supervised Learning is proposed for utilizing unlabeled data with the success of supervised learning. Producing a dataset with good labels is expensive, while unlabeled data is being generated all the time. The motivation of Self-Supervised Learning is to make use of the large amount of unlabeled data. The main idea of Self-Supervised Learning is to generate the labels from unlabeled data, according to the structure or characteristics of the data itself, and then train on this unsupervised data in a supervised manner. Self-Supervised Learning is wildly used in representation learning to make a model learn the latent features of the data. This technique is often employed in computer vision, video processing and robot control.
  • https://hoya012.github.io/blog/Self-Supervised-Learning-Overview/
  • https://paperswithcode.com/task/self-supervised-learning

ํ–‰๋ ฌ์˜ ๊ณฑ์€ linear transform.
bias๊ฐ€ ๋“ค์–ด๊ฐ€๋ฉด ํšŒ์ „.
๋น„์„ ํ˜•(activation function)์œผ๋กœ warp

vector-transform

sgd๋Š” lr์— ๋ฏผ๊ฐํ•˜๋‹ค.

Q & A

๋˜ํ•œ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰์ด ํด๋•Œ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ๋” ๋งŽ์ด ์ˆ˜์ง‘ํ•˜๋ฉด ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋œ๋‹ค๊ฐ€ ์ž˜ ์ดํ•ด๊ฐ€ ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์™œ ๋ฐ์ดํ„ฐ์˜ ๊ฐœ์ˆ˜๊ฐ€ ๋งŽ์œผ๋ฉด 12์ฐจ์˜ ๋ชจ๋ธ์ด 6์ฐจ์ฒ˜๋Ÿผ ๋ณด์ด๋Š”์ง€ ์ž˜ ์ดํ•ด๊ฐ€ ์•ˆ๋ฉ๋‹ˆ๋‹คใ…œใ…œ

  • ๋งค๋‹ˆํด๋“œ ๊ฐ€์ •
  • ์šฐ๋ฆฌ๊ฐ€ ์ƒ๊ฐ ํ•˜๋Š” ๋ฐ์ดํ„ฐ์˜ ํŒจํ„ด์€ ์ƒ๊ฐ๋ณด๋‹ค ์‹ฌํ”Œํ•˜๋‹ค.(์œ ์˜๋ฏธํ•œ ๊ฒƒ์€ ์ƒ๊ฐ๋ณด๋‹ค ์‹ฌํ”Œํ•˜๋‹ค)

ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๋งŽ๋‹ค.

  • ํ‘œํ˜„๋Šฅ๋ ฅ์ด ์ข‹์•„์ง.
  • ์ž์œ ๋„๊ฐ€ ๋†’์•„์งˆ์ˆ˜๋ก ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์ด ์ข‹์•„์ง.

๊ณต๊ฐ„์˜ ์ ‘ํžˆ๋Š” ๋ถ€๋ถ„์„ ๋ณด๋ฉด ๋Œ€์นญ์ด ๋˜๋Š” ์„ ์„ ๊ธฐ์ค€์œผ๋กœ๋งŒ ์ ‘๋Š”๊ฑฐ ๊ฐ™์€๋ฐ, ์ ‘ํž๋•Œ๋Š” ํ•ญ์ƒ ๊ทธ๋Ÿฐ์‹์œผ๋กœ๋งŒ ์ ‘ํžˆ๊ฒŒ ๋˜๋Š”๊ฑด๊ฐ€์š”?

  • ์ด์•ˆํŽ ๋กœ ์ฑ…์—์„œ ๊ทน๋‹จ์ ์œผ๋กœ ๋‚˜์˜จ๊ฒƒ์ž„.
  • Relu๋ฅผ ์ผ์„๋•Œ ๊ทธ๋žฌ์Œ.
  • ๋น„์Šทํ•œ๊ฒƒ๊ณผ ๋‹ค๋ฅธ๊ฒƒ์„ ๊ตฌ๋ถ„์„ ํฌ๊ฒŒํ•˜๊ธฐ ์œ„ํ•จ.

๊ณ„๋‹จํ•จ์ˆ˜๋Š” ํ™œ์„ฑํ•จ์ˆ˜ ์•„๋‹๊นŒ์š”โ€ฆ? ์†์‹คํ•จ์ˆ˜๋Š” (o-t) L2 or L1 or โ€ฆ๋กœ scalar๊ฐ’์œผ๋กœ ์•Œ๊ณ ์žˆ์Šต๋‹ˆ๋‹ค.

  • ๊ณ„๋‹จํ•จ์ˆ˜๋Š” ํ™œ์„ฑํ•จ์ˆ˜
  • ์†์‹คํ•จ์ˆ˜๋Š” scalar๊ฐ’.(๋ฒกํ„ฐ๋กœ ๋‚˜์™”์„ ๊ฒฝ์šฐ 1์ฐจ์›, 2์ฐจ์›, ์–ด๋Š ์ฐจ์›์œผ๋กœ ๊ธฐ์ค€์„?)

ํ”ผ์ณ๋งต์ด ์—ฌ๋Ÿฌ๊ฐœ์ด๋ฉด, L์ธต์ด ๋ณ‘๋ ฌ๋กœ ์—ฌ๋Ÿฌ๊ฐœ ์žˆ๋‹ค๊ณ  ๋ณด๋ฉด ๋˜๋‚˜์š”?? ์ธต์•ˆ ๋…ธ๋“œ๋“ค์ด ํ”ผ์ณ๋งต์˜ ๊ฐ๊ฐ์˜ ์š”์†Œ์ด๊ณ  ํ•œ ์ธต ์ž์ฒด๊ฐ€ ํ”ผ์ณ๋งต 1๊ฐœ๋ผ๊ณ  ์ดํ•ดํ•˜๊ณ  ์žˆ๋Š”๋ฐ ์ด๊ฒƒ์ด ๋งž๋Š”์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค.

  • ํ”ผ์ณ๋งต์€ cnn ํ•„ํ„ฐ์—์„œ ๋‚˜์˜จ ๊ฒƒ.
  • ๋ ˆ์ด์–ด ์ธต์„ ํ†ตํ•ด ๋‚˜์˜จ ์—ฌ๋Ÿฌ ํ”ผ์ณ๋งต์ด๋ผ๊ณ  ํ•ด๋„ ํ•˜๋‚˜์˜ ๋ฐ์ดํ„ฐ๋กœ ์ดํ•ด.

์„ ํ˜•๋Œ€์ˆ˜์™€ ํ™•๋ฅ ์„ ๊ธฐ์ดˆ๋ถ€ํ„ฐ ๊ณต๋ถ€ํ•˜๊ณ  ์‹ถ์€๋ฐ, ์ถ”์ฒœํ•˜์‹œ๋Š” ์ฑ…์ด๋‚˜ ๊ฐ•์˜๊ฐ€ ์žˆ์œผ์‹ ๊ฐ€์š”? ํ•œ๊ตญ์–ด๋กœ ๋˜์–ด์žˆ์œผ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹คโ€ฆ

  • ๊ธฐ๋ณธ์ ์ธ ๊ฐœ๋…?
  • ๋”ฅ๋Ÿฌ๋‹์— rank๋Š” ๋ณ„๋กœ ํ•„์š”ํ•˜์ง€ ์•Š์Œ..
  • ๊ทธ๋•Œ๊ทธ๋•Œ ๊ตฌ๊ธ€๋ง.

๋‹จ์ธตํผ์…‰ํŠธ๋ก ์—์„œ์˜ b ๋Š” ์ž„๊ณ„๊ฐ’์˜ ์—ญํ• ์„ ํ•œ๋‹ค๊ณ  ๊ฐ•์˜์—์„œ ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด ๋‹ค์ธต ํผ์…‰ํŠธ๋ก , deep learning ์—์„œ์˜ ๊ฐ๊ฐ์˜ b1, b2 ๋“ฑ.. ์˜ ๊ฐ’๋“ค๋„ ๊ฐ ํผ์…‰ํŠธ๋ก ์—์„œ์˜ ์ž„๊ณ„๊ฐ’ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ธ๊ฐ€์š”??

  • b๊ฐ’์€ ์ž„๊ณ„๊ฐ’์€ ๋งž์Œ. ์ง€๊ธˆ์€ ์ž„๊ณ„๊ฐ’์˜ ๊ธฐ์ค€์„ 0์œผ๋กœ ๋งž์ถฐ๋†“๊ณ  bias term์œผ๋กœ ๋บ์Œ.
  • ์š”์ฆ˜์— ๋‚˜์˜ค๋Š” ์‹ ๊ฒฝ๋ง์˜ activation function์˜ ๊ธฐ์ค€์ ์€ 0์ž„. ์น˜์šฐ์ง„๊ฑธ 0์œผ๋กœ ์˜ฎ๊ฒจ๋†“๊ธฐ ์œ„ํ•ด bias term์„ ํ•˜๋‚˜ ๋” ์คŒ.

ReLU ํ•จ์ˆ˜๋ฅผ ์ด์šฉํ•ด์„œ ๋น„์„ ํ˜• ๊ณก์„  ํ•จ์ˆ˜๋ฅผ ๊ทผ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ทผ๊ฑฐ๊ฐ€ ๋ญ”์ง€ ์•Œ๊ณ ์‹ถ์Šต๋‹ˆ๋‹ค

  • ์ด๋ก ์ ์ธ ๋ถ€๋ถ„

์ฐจ์ˆ˜๋Š” ๋‹คํ•ญ์‹์˜ ๋ชจ๋ธ์„ ์“ฐ๋Š” ๊ฒฝ์šฐ์— ์ฐจ์ˆ˜๊ฐ€ ๋งŽ์•„์ง€๋ฉด ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ๋งŽ์•„์ง€๋Š” ๊ฒƒ.

nips, cvpr ๋…ผ๋ฌธ์„ ์ง„ํ–‰ํ•ด๋ณด๋Š” ๊ฒƒ๋„..?

Leave a comment