ML Ethics, Fairness, and Inclusiveness

Machine learning ethics refers to the moral principles and guidelines that govern the development and deployment of AI systems. Fairness ensures that AI models do not discriminate against individuals or groups based on sensitive attributes like race, gender, age, or socioeconomic status. Inclusiveness means that AI systems are designed to benefit all people, including those with disabilities or from underrepresented groups. These concepts are intertwined: an unfair model is unethical, and an exclusive design can lead to unfair outcomes.

Microsoft has established six Responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. The AI-900 exam expects you to understand these principles and how they apply to common AI workloads. You must be able to identify potential sources of bias in data and models, and know the tools Azure provides to detect and mitigate bias.

Bias can enter an AI system at multiple stages: - Data bias: The training data may not represent the full population. For example, a facial recognition system trained mostly on light-skinned faces will perform poorly on darker skin tones. This is called *sampling bias*. - Label bias: Human annotators may introduce their own biases when labeling data. For instance, if a sentiment analysis dataset is labeled by only one demographic, the labels may reflect that group's perspective. - Algorithmic bias: The model's architecture or optimization process may favor certain outcomes. For example, a model that minimizes overall error might perform well on the majority group but poorly on minorities. - Deployment bias: Even a fair model can become biased if deployed in a context different from its training environment. For instance, a model trained on urban data may fail in rural areas.

To quantify fairness, several metrics exist. The AI-900 exam focuses on the concept of *fairness* rather than specific math, but knowing these metrics helps: - Demographic parity: The proportion of positive outcomes should be equal across groups. - Equal opportunity: The true positive rate should be equal across groups. - Equalized odds: Both true positive and false positive rates should be equal across groups.

Azure Machine Learning provides the Fairness Dashboard to compute these metrics and compare model performance across sensitive groups. You can specify a sensitive feature (e.g., race, gender) and the dashboard will show disparities.

Azure provides several tools to implement these principles: - Azure Machine Learning Fairness Dashboard: Visualizes fairness metrics across groups. You can upload a model, specify sensitive features, and see disparities. - InterpretML: An open-source package that explains model predictions. It includes glassbox models (e.g., Decision Trees) and blackbox explainers (e.g., SHAP). - Error Analysis Dashboard: Identifies where the model makes the most errors, often revealing bias in specific subgroups. - Counterfactual Analysis: Shows how changing a feature (e.g., age) would change the prediction, helping to understand bias. - Data sheets: Microsoft encourages creating data sheets that document the dataset's provenance, demographics, and potential biases.

Mitigation can occur at three stages: - Pre-processing: Adjust the training data to reduce bias. Techniques include reweighting samples, resampling to balance groups, or removing sensitive features. However, removing sensitive features may not eliminate bias if other features correlate with them. - In-processing: Modify the learning algorithm to penalize unfairness. For example, adding a fairness constraint to the loss function. - Post-processing: Adjust the model's outputs to achieve fairness. For instance, changing decision thresholds for different groups to equalize outcomes.

Inclusiveness goes beyond fairness. It means designing AI that works for everyone, including people with disabilities. Microsoft's inclusive design principles include: - Recognize exclusion: Understand how your AI might exclude certain users. - Learn from diversity: Involve diverse users in the design process. - Solve for one, extend to many: Designing for one person with a disability often benefits everyone.

Examples: Speech-to-text that works for non-native speakers, computer vision that recognizes diverse body types, and chatbots that support multiple languages.

Transparency means users should know they are interacting with an AI and understand how decisions are made. Azure provides: - Model Interpretability: Use InterpretML to get global and local feature importance. - Explainable AI (XAI): Techniques like SHAP and LIME explain individual predictions. - Documentation: Model cards and data sheets provide standardized documentation.

Organizations must establish governance for AI. This includes: - An AI ethics board to review high-risk applications. - Audit trails to track model versions and decisions. - Impact assessments to evaluate potential harms.

ML ethics, fairness, and inclusiveness are foundational to responsible AI. Microsoft's six principles guide the development of ethical AI. Azure provides tools to detect and mitigate bias, ensure transparency, and promote inclusiveness. For the AI-900 exam, focus on understanding the principles, common bias sources, and the purpose of each Azure tool.

A major bank uses an AI model to approve personal loans. The training data includes historical decisions from the past 10 years, which reflect societal biases: women and minorities were disproportionately denied loans. The model learns to replicate these patterns. When deployed, it rejects loan applications from qualified women at higher rates. The bank uses Azure Machine Learning's Fairness Dashboard to detect that gender and race are causing disparities. They apply a pre-processing technique called 'reweighing' to assign higher weights to underrepresented groups in the training data. After retraining, the model's demographic parity improves. However, the bank must also ensure it doesn't violate anti-discrimination laws. They document the model in a model card and set up monitoring to detect drift. The challenge is balancing fairness with profitability; overly aggressive fairness constraints could lead to higher default rates.

A hospital deploys an AI system to diagnose skin cancer from images. The training dataset is mostly from Caucasian patients, so the model performs poorly on darker skin tones. This is a safety issue because misdiagnosis can be fatal. The hospital uses Azure's Error Analysis Dashboard to find that error rates are highest for patients with dark skin. They collect more diverse images and retrain the model. They also use data augmentation to simulate different skin tones. The model's accuracy improves across all groups. The hospital also implements transparency by providing explanations for each diagnosis. They use InterpretML to show which features (e.g., asymmetry, border irregularity) influenced the decision. This helps doctors trust the AI. The project highlights the need for inclusive data collection and continuous monitoring.

A tech company uses an AI to screen résumés. The training data is from past hires, which are predominantly male. The model learns to favor male candidates. To fix this, the company removes gender-related features and uses a fairness constraint during training. They also use counterfactual analysis to see if changing the gender in a résumé changes the prediction. They find that even after removing gender, features like 'years of experience' correlate with gender because women often have career gaps. They apply a post-processing technique that equalizes the selection rate across genders. The company also ensures transparency by explaining to candidates how the AI works. They publish a model card and allow candidates to request a manual review. This builds trust and reduces legal risk.