Responsible ML Daron Yöndem http://daron.me @daronyondem

The What and the Why? • Fairness: ML models may behave unfairly by negatively impacting groups of people, such as those defined in terms of race, gender, or age. • Interpretability: Ability to explain what parameters are used and how the models “think” to explain the outcome for regulatory oversight. • Differential Privacy: Monitoring applications’ use of personal data without accessing or knowing the identities of individuals

No content

No content

No content

No content

Fairlearn • Fairness Assessment • Fairness Mitigation (in Classification and Regression Models) • During or after model building. • Open Source https://github.com/fairlearn • Integrated into Azure Machine Learning

Fairlearn Demo

InterpretML • Glassbox Models (Decision trees, rule lists, linear models, Explainable Boosting Machine) • Blackbox Models (Existing model) • Explanations are approximations • Open Source https://github.com/interpretml

InterpretML Demo

Differential Privacy The guarantee of a differentially private algorithm is that its behavior hardly changes when a single individual joins or leaves the dataset. • Smart Noise https://smartnoise.org/ This toolkit uses state-of-the-art differential privacy (DP) techniques to inject noise into data, to prevent disclosure of sensitive information and manage exposure risk.

Links worth sharing Microsoft Learn : Explore differential privacy • https://drn.fyi/2QRL3V1 Capgemini “AI and the ethical conundrum” Report • https://drn.fyi/3gBsz5G IDC report: Empowering your organization with Responsible AI • https://drn.fyi/3sKilCx

Thanks http://daron.me | @daronyondem Download slides here; http://daron.me/decks