Group Study (2020-2021)/Deep Learning

[DeepSleep] ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ ์Šคํ„ฐ๋”” 1์ฃผ์ฐจ

morijwana 2021. 7. 11. 22:51

๐Ÿ’ ํ•˜๋žŒ: https://github.com/dsc-sookmyung/2021-DeepSleep-Paper-Review/blob/main/Week1/resnet.md

  • ์ฃผ์ œ: Deep Residual Learning for Image Recognition (2015, ResNet)
  • ๋งํฌ : https://arxiv.org/pdf/1512.03385.pdf
  • ๊ฐœ์š”/๋ฐฐ๊ฒฝ
    • Batchnorm ๋„์ž… ์ดํ›„์— 10๊ฐœ ์ด์ƒ์˜ ๋ ˆ์ด์–ด๋ฅผ ํ•™์Šต์‹œํ‚ฌ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค.
    • ์ธต์ด ๋งŽ์€ ๋ชจ๋ธ์ผ์ˆ˜๋ก ์ธต์ด ์ ์€ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•˜๊ฑฐ๋‚˜ ๋” ์ข‹์€ ์„ฑ๋Šฅ์ด ๋‚˜์™€์•ผ ํ•œ๋‹ค๊ณ  ๊ธฐ๋Œ€ํ•˜์˜€๋‹ค
    • ํ•˜์ง€๋งŒ ์—ฌ์ „ํžˆ ๊นŠ์€ ๋ชจ๋ธ์€ ์–•์€ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ข‹์ง€ ์•Š์•˜๋‹ค.
    • ์ผ๋ถ€ ๋ ˆ์ด์–ด๊ฐ€ identity function (f(x)=x) ๋ฅผ ์ž˜ ๋ฐฐ์šธ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋ฉด ์–ด๋–จ๊นŒ?
  • ๋ชฉํ‘œ : ๋ชจ๋ธ์ด identity function ์„ ๋” ์‰ฝ๊ฒŒ ๋ฐฐ์šธ ์ˆ˜ ์žˆ๋„๋ก ๋„คํŠธ์›Œํฌ๋ฅผ ์กฐ์ ˆํ•˜์ž!
  • ์ฃผ์š” ์šฉ์–ด
    • Plain block / Residual Block
    • StagesResidual Networks

 

๐Ÿ’ ์ˆ˜์—ฐ: https://github.com/dsc-sookmyung/2021-DeepSleep-Paper-Review/blob/main/Week1/AlexNet.md

  • ์ฃผ์ œ: Imagenet classification with deep convolutional neural networks (AlexNet)
  • ๋งํฌ: http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
  • ๋ฐฐ๊ฒฝ
    • ํ˜„์žฌ MNIST ๋ถ„๋ฅ˜๊ธฐ์˜ ์ธ์‹ ๋Šฅ๋ ฅ(์—๋Ÿฌ ๋ฐœ์ƒ๋ฅ  0.3% ์ดํ•˜)์€ ์ธ๊ฐ„์˜ ์„ฑ๋Šฅ๋งŒํผ ๋›ฐ์–ด๋‚˜์ง€๋งŒ, ํ˜„์‹ค์˜ ์‚ฌ๋ฌผ๋“ค์€ ํ›จ์”ฌ ๋” ๋‹ค์–‘์„ฑ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฏ€๋กœ ๋” ํฐ ํฌ๊ธฐ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”
    • LabelMe, ImageNet ๋“ฑ ๋งค์šฐ ํฐ ๊ทœ๋ชจ์˜ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์˜ ๋“ฑ์žฅ
    • ์ด๋ฏธ์ง€์˜ ํŠน์„ฑ์— ๋Œ€ํ•ด ๊ฐ•๋ ฅํ•œ ์ถ”๋ก ์„ ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ํ›ˆ๋ จํ•˜๊ธฐ๋„ ์‰ฌ์šด CNN์„ ์‚ฌ์šฉ (๋‹ค๋งŒ ์„ฑ๋Šฅ์€ standard feedforward NN๋ณด๋‹ค ์กฐ๊ธˆ ๋‚˜์  ์ˆ˜ ์žˆ์Œ)
    • ILSVRC-2010๊ณผ ILSVRC-2012์—์„œ CNN์„ ์ด์šฉํ•œ ๋ชจ๋ธ(AlexNet)๋กœ ์šฐ์Šน
  • ์ฃผ์š” ๋‚ด์šฉ
    • AlexNet์˜ ๋„คํŠธ์›Œํฌ ์„ฑ๋Šฅ ํ–ฅ์ƒ๊ณผ ํ›ˆ๋ จ ์†๋„ ๋‹จ์ถ• ๋ฐฉ๋ฒ•
    • AlexNet์˜ Overfitting์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋ฒ•
    • AlexNet์˜ ๊ตฌ์กฐ