Explainable AI

Interpreting Neural Network (ntu.edu.tw)

  • Interpretable v.s. Powerful 

    • • Are there some models interpretable and powerful at the same time? • How about decision tree? (A tree can still be terrible! We use a forest!)

  • Goal of Explainable ML 

  • This is a “cat”.

    • Local Explanation —— What does a “cat” look like?
    • Global Explanation—— Why do you think this image is a cat?

  • Local Explanation: Explain the Decision

    • Which component is critical?——Removing or modifying the components Important component • Large decision change 没它不行（用灰色方块遮住 在不同位置）

    • 图片

    • How a network processes the input data？

      • visualization图片

      • probing图片

  • GLOBAL EXPLANATION: EXPLAIN THE WHOLE MODEL

    • 图片
      图片
      图片
      图片

图片

Domain Adaptation

Domain shift: Training and testing data have different distributions

  • Little but labeled 图片

  • Large amount of unlabeled data

    • 图片

    • 让source（）labeled和target（unlabeled）分不出差异图片

    • 图片

    • 可能变成只是为了分开而分开 但分类不正确