Why Deep?

比較deep和shallow model

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 比較有著一樣多參數的deep和shallow model,可以發現deep的performance較好

deep model比較好,為什麼?

1.情境1

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 從圖中可以發現長頭髮的男生data比較少,造成detect長髮男的performance較差

2.情境2

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 若是換個方式,把原來的問題切成比較小的問題,去learn一些classifiers去detect某種特徵地出現。
  • 以上圖為例,雖然長髮男生少,但男生和女生的資料足夠多。雖然長髮男生少,但長髮和短髮的data是足夠多的。

3.Modularization 模組化

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 每個classifier會去參考到basic classifier的結果

4.Deep與模組化

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • deep的第一層為最basic的classifier,而後的每層layer會把前一層當成input的module
  • 而deep learning中做出模組化是machine自動學從data到的
  • 當deep是模組化,把問題變簡單了,可以發現需要的train data是需要比較少的,因此deep learning並不是big data

語音

  • deep learning在影像和語音上表現特別好

過去的HMM-GMM

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 使用GMM描述分布的機率
  • 問題:tri-phone太多了,每個state也都要用一個GMM描述(也有部分共用的方法)
  • 沒有效的原因:所有的phone或state會被當作independent,但實際上它們之間是有相關性的,不是independent的

DNN(deep learning)

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 所有state都共用一個DNN
  • 能有效率的使用參數(由於發音的特性)

End-to-end Learning

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

  • 只給model input、output讓他自己學習每個步驟
  • 當我們要處理一個很複雜的問題時,就可以使用deep learning,疊一個很深的network去做分工學習

圖像化比喻Deep learning

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

tags: ML2020