Fair and Transparent Machine Learning Methods (FTML)

Machine learning systems are increasingly shaping important aspects of our daily lives from the content we see online and the job opportunities we're offered, to healthcare recommendations and financial services we can access. While these systems can be remarkably effective, they also raise important questions. How can we assess if a model treats different demographic groups equitably, and intervene when it doesn't? How do we prevent biases in the training data, such as historical inequalities, from perpetuating these patterns in future uses? And how do we improve transparency in applications necessitating an understanding of the reasoning behind decisions?

This course addresses these questions by focusing on the methods to assess and improve the fairness, accountability and transparency of machine learning models. As such, it assumes students have prior knowledge of machine learning.

This course covers the following tentative topic list:

• Understanding and documenting the responsible ML lifecycle: datasheets, model cards, model misuse, and challenges with open model deployment.
• Statistical notions and metrics for fairness and bias.
• Mitigating bias and learning fair representations: counterfactual data augmentation, adversarial debiasing, and model calibration.
• Mechanistic model interpretability: probing neural networks, neuron and circuit analysis.
• Generating and evaluating explanations: post-hoc methods (e.g., SHAP), natural language explanations (e.g. Chain-of-Thought), and assessing explanation quality and faithfulness.
• Advanced topics: navigating fairness-accuracy trade-offs, privacy-preserving approaches to fair ML, and emerging challenges in responsible AI.