Try   HackMD

Semi-supervised Learning 半監督式學習

intro

  • semi-supervised learning 由少部分labeled data和大部分unlabeled data構成
    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 →
  • 藉由資料集的分布,或模型的判斷,來自行標示資料。

實作

  • Semi-supervised Generative Model
    不斷加入unlabeled data,影響原本generative model的分布。
    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 →
    推導

    每加入一筆資料,重新估測一次分布模型,影響的程度由data屬於那個class的機率決定(越有可能是A-Class,對A的影響越大)。

  • Self-training
    由現有模型判斷機率,如果屬於機率A比較大,將那筆資料視為A資料,再放回訓練。
    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 →
  • Entropy-based Regularization
    如果資料集分布很明確(Ex.不是狗就是貓),通過最小化loss function,要求Entropy降低(Entropy越低,資料越集中),達成正規化。
    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 →
  • Cluster and then label
    劃分為同一cluster的unlabeled data,視為那一個cluster的label。需要能明確劃分,才能得到比較好的效果。
    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 →
  • Graph-based Approach
    圖論中,相連接的Data,視為同一個label,並且會沿著edge影響其他data。
    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 →
    推導
  • Better Representation
    尋找更好的因素來區分資料。Ex.想要找鬍子,先找到頭??
    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