Fair and Transparent Machine Learning Methods (FTML)

Deploying machine learning models for downstream applications brings with it a wealth of possibilities. However, there is also a non-negligible risk of potential harm if models are not developed carefully.

Data can encode undesired societal biases, which can in turn be perpetuated by machine learning models if trained on such data. There may be risks in developing automated solutions for certain application tasks altogether. Moreover, ML models are often black boxes whose decisions are not transparent to end-users, creating imbalances and issues regarding the accountability of models. Therefore, it is imperative to reflect on the benefits and risks of ML models, to develop methods to detect and mitigate biases in ML models, and to create solutions to make the inner workings of models more transparent. This course focuses on the technical solutions needed to 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:

• Statistical notions of fairness and bias
• The intended usage of ML models, e.g. datasheets, model cards
• Learning fair representations, e.g. counterfactual data augmentation, adversarial training, model calibration
• Model interpretability and transparency
• Generating explanations, e.g. post-hoc explainability, generating free-text explanations
• Evaluating model explanations
• Probing representations for bias, e.g. functional testing, subspace probing, generative approaches