WHAT: Danielle Villa on "Honesty Beyond Fact Checking - AI Faithfulness"
WHEN: Weds, 10 Apr (6p)
IN-PERSON: Winslow 1140
EVENT PAGE: https://bit.ly/4cUkrrJ
DESCRIPTION: Explanations for AI decisions, especially LLMs, have become especially prominent in research towards trustworthy AI. However, just because an explanation sounds reasonable does not mean that it accurately reflects the internal decision process of the model. Rather than just ensuring that explanations be factually accurate and logically sound, many call for explanations to be faithful as well. In this talk we'll discuss the difficulty in measuring faithfulness in explanations, existing explanation faithfulness metrics, and whether or not faithfulness is something that can ever be guaranteed.
BIO: Danielle is a 2nd year PhD student studying how semantic technologies can be used to improve LLMs, with a focus on improving evaluation metrics using knowledge graphs. Her current work is on a process for generating counterfactuals to question answering datasets using knowledge graphs to evaluate the faithfulness of LLM-generated explanations.