ML Ethics — Responsible AI

AI systems must be fair, transparent, and accountable. Ethics is not optional — it's a legal requirement in many jurisdictions.

Fairness

Types of fairness:
├─ Demographic parity: Equal positive rates across groups
├─ Equal opportunity: Equal true positive rates
├─ Equalized odds: Equal TPR and FPR
└─ Individual fairness: Similar individuals treated similarly

Sources of bias:
├─ Historical bias in training data
├─ Representation bias (underrepresented groups)
├─ Measurement bias (proxy variables)
└─ Aggregation bias (one model for diverse groups)

Interpretability

SHAP (SHapley Additive exPlanations):
├─ Game theory-based
├─ Assigns contribution to each feature
└─ Consistent and locally accurate

LIME (Local Interpretable Model-agnostic Explanations):
├─ Approximate model locally with interpretable model
├─ Feature importance per prediction
└─ Model-agnostic

Key Takeaways

1. Fairness requires measuring and mitigating bias
2. SHAP and LIME provide model interpretability
3. Privacy requires differential privacy and federated learning
4. Transparency means documenting models and data
5. Accountability requires human oversight
6. Ethics is a legal requirement (EU AI Act)
7. Diverse teams reduce blind spots
8. Regular audits catch emerging issues