Oana-Maria Camburu (University College London)
Oana-Maria Camburu is a Senior Research Fellow with a Leverhulme Early Career Fellowship, University College London. She works mainly on Explainable AI, building neural networks that generate human-like explanations for their predictions. She did her Ph.D. on "Explaining Deep Neural Networks", working with Prof. Phil Blunsom and Prof. Thomas Lukasiewicz at the Department of Computer Science, University of Oxford.
Title: Can AI Models Give Us Correct and Faithful Natural Language Explanations for Their Predictions? (Slides)
Abstract: As the current ground-breaking advancements are making AI models increasingly more present in our lives, having human-interpretable explanations for their predictions is not only of regulatory importance in high-risk domains but also a way to generally improve the trustworthiness of the models. To this end, we argue that explanations should be at least correct and faithful. First, an explanation is said to be correct if it provides a ground-truth justification for solving the task. Models with incorrect explanations would likely be of little usage for end users. Second, an explanation is said to be faithful if it reflects the internal decision-making process of the model. Unfaithful explanations would be "lying" to the users and may dangerously influence their perception of the model and, in turn, their decisions. In this talk, we will see a series of models, benchmarks, adversarial attacks, and metrics to examine the capabilities of AI models to correctly and faithfully explain their predictions in natural language.
Abstract: As the current ground-breaking advancements are making AI models increasingly more present in our lives, having human-interpretable explanations for their predictions is not only of regulatory importance in high-risk domains but also a way to generally improve the trustworthiness of the models. To this end, we argue that explanations should be at least correct and faithful. First, an explanation is said to be correct if it provides a ground-truth justification for solving the task. Models with incorrect explanations would likely be of little usage for end users. Second, an explanation is said to be faithful if it reflects the internal decision-making process of the model. Unfaithful explanations would be "lying" to the users and may dangerously influence their perception of the model and, in turn, their decisions. In this talk, we will see a series of models, benchmarks, adversarial attacks, and metrics to examine the capabilities of AI models to correctly and faithfully explain their predictions in natural language.
Michael Yeomans is an Assistant Professor of Organisational Behaviour at Imperial College Business School. His research focuses on how machine learning can be used to understand and improve human judgment and decision-making - in particular, by using natural language processing to study decisions during conversations. His research has been published in many journals, including Journal of Personality and Social Psychology, Proceedings of the National Academy of Sciences, Management Science, and Organizational Behavior and Human Decision Processes. He received his Ph.D. from the University of Chicago and prior to joining Imperial, he completed a post-doctoral fellowship at Harvard University.
