AIR Seminar with Fuxin Li | Hariri Institute | Boston University

Value of the Information & Strength of the Argument

The talk provides valuable insights into the state of explainable AI, presenting original research with concrete evidence and examples. Li argues convincingly that traditional gradient-based saliency maps are unreliable and demonstrates how his methods outperform baseline approaches. He supports his claims with visualizations and quantitative metrics, building a coherent narrative from motivation to methodology. The discussion on feature polysemanticity and the importance of staying in distribution is particularly illuminating. However, the argumentation relies heavily on the presenter’s own work without external validation, and some claims lack direct comparison with recent alternative methods beyond a few baseline comparisons.

Scientific Rigor, Source Quality, Title Accuracy

The seminar is rigorous, referencing established work like that of Kate Saenko’s group and citing the DARPA XAI program. Li explains the theoretical basis for his methods and acknowledges limitations. The title accurately matches the content, and the description provides relevant context. The talk mentions several awards and publications, indicating a strong research foundation. However, no external sources or references are provided beyond the seminar link, and the talk is a single presentation without peer review. The audience questions are addressed, but the content remains mostly one perspective.

The title accurately describes the seminar format and the speaker, and the content aligns fully with the talk abstract.

8/10

The speaker, Prof. Fuxin Li, is a recognized researcher in explainable AI with multiple awards, and the talk presents original work with clear methodology, though lacks published data or external peer review during the seminar.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction of Fuxin Li by host, including his background and awards.
• 03:00Motivation for explainability in deep learning: debugging failures and safety.
• 10:00Failed attempt to summarize network features; feature polysemanticity discovery.
• 20:00Introduction of deletion-insertion metrics and the I-GOS algorithm for heatmaps.
• 30:00IGOS++ improvements for high-resolution heatmaps using bilateral filtering.
• 40:00Structural attention graphs (SAG) for visualizing multiple decision paths.
• 50:00Counterfactual explanations using VAEs and GANs to transform images.

• Hariri Institute Events — Link in video description for more information about Hariri Institute seminars.

• Hariri Institute Events — Official event page, provides context for the seminar.

• No conflicting sources found — The talk does not explicitly contradict other known sources, but the claims on heatmap reliability may challenge some prior saliency methods.