Manya Singh

This project addresses responsible AI by studying robust machine learning models designed to refrain from predictions when uncertain, providing transparent explanations for the abstention. The purpose of these explanations is to a) enhance user comprehension and b) empower domain experts to pinpoint inaccuracies for model refinement. As the "uncertainty" part can be a result of noisy data, outliers, overlapping decision boundaries, etc. are subjective to the domain, therefore ideally this project will validate the outcome of the explanation using user studies. Our focus involves assessing current algorithms, adapting them to this context, or innovating new ones to align with our objectives. Rigorous testing across different domains, ideally incorporating user testing, will gauge the efficacy of the explanations, ensuring the models meet both transparency and functionality standards.