- Learn
- Machine Learning
- Responsible ML
ML-M16
Responsible ML
Building models that are fair, explainable, private and robust: measuring bias, explaining predictions, protecting training data, and surviving adversarial inputs and shifting data.
A model that is accurate on average can still deny loans unfairly to one group, refuse to say why, leak the medical records it was trained on, or fall over when someone changes a few pixels. None of these failures show up in a single test-set score, and all of them have ended real projects.
You will start with fairness: where bias enters a system, how competing fairness metrics are defined, and why they cannot all be satisfied at once. Then you will explain models globally and one prediction at a time, protect training data with differential privacy and federated learning, and finish with adversarial examples, distribution shift and how to test for robustness before users do it for you.
Lessons
- ML-16.01 Fairness and bias Models can reproduce and amplify biases present in historical data. Intermediate 12 min
- ML-16.02 Explainability Explainability helps people trust, debug and challenge models. Intermediate 10 min
- ML-16.03 Privacy-preserving ML Models can leak information about training data. Advanced 12 min
- ML-16.04 Robustness and adversarial examples Tiny, carefully crafted changes to an input can make a model confidently wrong; these are adversarial examples. Advanced 10 min