PhD Dissertation Proposal by ZHANG Mengdi | Towards Explainable Neural Network Fairness

Neural networks are achieving higher accuracy and exceptional performance and widely applied in solving many real-world problems. At the same time, they are shown to be vulnerable to attacks, difficult to debug and subject to fairness issues. To improve people's trust in the technology, it is necessary to provide some human-understandable explanation of neural networks' decisions, e.g., why is that my loan application is rejected whereas hers is approved? It becomes a fairness issue if the only explanation possible is based on certain sensitive features (such as race and gender). In this thesis, we develop multiple approaches towards improving people's trust on neural networks.

We propose measurements on the decision explainability of neural networks. Afterwards, we develop multiple algorithms that allow us to automatically evaluate and compare neural networks in terms of the measurements. We propose \sgde, an interpretable testing approach which systematically identifies and measures hidden (which we call `subtle') group discrimination of a neural network characterized by conditions over combinations of the sensitive features. We propose an approach which adaptively chooses the fairness improving method based on causality analysis.

While most existing fairness improvement methods have shown effectiveness on tabular data, we are still working on textual data and mitigating bias on NLP models such as BERT.

Mengdi ZHANG is a Ph.D. candidate at SMU SCIS, supervised by Prof. Jun SUN. Her research interests are mainly on AI security, including machine learning interpretability, fairness testing and discrimination mitigation.