Explainable AI: Demystifying Complex Models with Shapley Values - Neeraj Pandey

歡迎來到 PyCon TW 2023 共筆

共筆入口：https://hackmd.io/@pycontw/2023
手機版請點選上方 按鈕展開議程列表。
Welcome to PyCon TW 2023 Collaborative Writing
Collaborative Writing Workplace：https://hackmd.io/@pycontw/2023
Using mobile please tap to unfold the agenda.

Collaborative writing start from below
從這裡開始共筆

Here is the link to the tutorial and slides
Slides

Explainable AI: Help human understand how AI models make certain prediction. Unveil the black box model
e.g. Doctors need to understand why a model makes a certain prediction to trust it and to effectively communicate the risk to patients

• The black box dilemma
  • Simple model is easier to interpret (e.g. linear regression model). Although complicated neural networks are powerful, they are too complex to interpret.

XAI methods

• LIME: local interpretation
• SHAP

Much accuracy, less interpretability

Benefit

• Model stability

  • Tolereance for error
  • Uniform behavior

• Cause and effect

  • Beyond correlation
    Understanding causality is important
  • Informed Predictions

• User confidence

  • AI-Human Synergy
  • Openness as Trust builder

• lawful accountability

  • Right to explanation
  • Accountability
    • Who is responsible for AI decision?

Types of explainable AI methods

Types of XAI methods

• Model specific
• Model agnostic

Model scope

• Global
  Understand the overall behavior of the model
• Local
  explain why the model made a specific prediction for a given data point

Global explainability

• feature importance
  Like a permutation feature importance
• partial dependence plots (PDPs)
  show how individual features affect predictions across different data points.
• linear or logistic regression with regularization

Local explainability

• LIME
• Conterfactual explanations
• SHAP

Below is the part that speaker updated the talk/tutorial after speech
講者於演講後有更新或勘誤投影片的部份