AARC Seminar: Debugging and Explaining Unfairness in Machine Learning Models

The monthly AARC Seminar will feature speakers actively researching applied AI from both industry and academia. This seminar series is open to all Purdue faculty, staff, postdocs, and students.

For the 2024-25 academic year, seminars will take place on Fridays at 1:00PM in A. A. Potter Engineering Center (POTR), room 234. Details, including talk titles, will be posted here as they are available.

Algorithmic decision-making systems continue to garner concerns regarding the perpetuation of systemic biases reflected in their training data. Discriminatory outcomes violate human rights and undermine public trust in automated decision-making. This talk will describe our efforts toward rendering AI-based systems less biased. We present Gopher, a causal databased explanation mechanism, that allows us to diagnose the outcomes of a machine learning model and detect subsets of the training data most responsible for biased decisions. Gopher quantifies and approximates the causal responsibility of subsets and prunes the huge search space to generate compact, interpretable, and causal explanations for biased model predictions.