Increase trust in AI systems

Just like medical treatments or economic policies, AI systems can be considered as interventions in our society. I build methods to help experts interpret an AI system’s output and estimate their (differential) impact on different subpopulations. One of these methods, Diverse Counterfactual Explanations (DiCE), has been integrated as a part of Microsoft’s Responsible AI platform.

I also work on improving the generalizability of ML models so that they are less sensitive to their training data distribution. This line of work has led to a NASSCOM AI GameChangers Award (2023-24).

Counterfactual explanations

• A methods that generates diverse counterfactual explanations for explaining ML classifiers (Mothilal et al., 2020)
• Using LLMs-generated counterfactuals to explain AI model’s predictions (Gat et al., 2024)
• Unifying feature attribution and counterfactual explanation methods: When to use which (Kommiya Mothilal et al., 2021)
• Evaluating and mitigating bias in image classifiers (Dash et al., 2020)

Building prediction models that generalize better

• The challenges of prediction and explanation in social systems (Hofman et al., 2017)
• Using causal reasoning to build generalizable ML classifiers (missing reference)

Role of causality in trustworthy ML

• The necessary role of causality in understanding when ML explanations can improve human understanding (Chen et al., 2023)
• A framework for deploying trustworthy ML systems based on technology readiness levels (Lavin et al., 2022)
• ML fairness estimates can be misleading without modeling data missingness (Goel et al., 2021)