HCII Seminar Series - Polo Chau
Speaker
Polo Chau
Professor at Georgia Tech, School of Computational Science and Engineering
When
-
Where
Newell-Simon Hall 1305
Video
Video link
Description
"Visual and Algorithmic Interpretation for Responsible AI"
Our group, the Polo Club of Data Science, develops novel visual and algorithmic tools to help make sense of AI behaviors and risks. Our Safe AI research thrust investigates AI vulnerabilities and develops countermeasures to strengthen AI safety. We pioneer LLM safety landscape visualization, revealing a novel "safety basin" phenomenon observed universally in LLMs that highlights the critical role of system prompts in protecting models from misuse. To promote safer LLM fine-tuning, we introduce dynamic safety shaping (DSS), the state-of-the-art technique that identifies safe and unsafe text segments to preserve alignment during fine-tuning. LLM Self Defense presents a straightforward and practical method for detecting harmful outputs through self-examination.
Our complementary interpretable AI research creates a suite of interactive visualizations that enhance people's ability to understand complex models and their vulnerabilities. WizMap enables scalable, on-device exploration of large AI embeddings, while LLM Attributor introduces a new way to quickly attribute an LLM’s text generation to specific training data points, aiding in the inspection of model behavior. Our state-of-the-art ConceptAttention algorithm (top 1%, ICML'25) visualizes any text concepts in generated images and videos without requiring model training, while the CRAYON algorithm corrects inaccurate model attention using simple yes–no human feedback, evaluated with nearly 6,000 participants.
Our AI Explainers (Transformer Explainer, Diffusion Explainer, CNN Explainer, GAN Lab, ManimML), used by more than 1 million people across 200+ countries, provide accessible tools for both students and experts to learn about AI models.
Speaker's Bio
Duen Horng (Polo) Chau is a Professor at Georgia Tech’s School of Computational Science and Engineering, holding a Machine Learning (ML) Ph.D. from Carnegie Mellon University (Dissertation Award, Honorable Mention). He co-directs the MS Analytics program. He was the Director of Industry Relations at The Institute for Data Engineering and Science, and the Associate Director of Corporate Relations at the Center for Machine Learning. His research bridges ML and visualization, creating scalable interpretable tools that enhance human understanding of large-scale data and complex ML models. He received 19 best paper/poster/demo type awards and published 200+ refereed articles. He received many faculty awards (Google, Meta, Intel), the GT-wide Senior Faculty Outstanding Undergraduate Research Mentor Award, and the Dean’s Award for Faculty (group) for establishing GT’s AI leadership for the College of Computing. His work has been deployed by Google, Microsoft, Meta, Nvidia, NASA, ADP, NortonLifeLock, and Atlanta Fire Rescue Department. His students won PhD fellowships (Google, Apple, IBM, JPMorgan, NASA, NSF). He teaches 1,000+ students each semester. His work has been covered by popular media: The Wall Street Journal, Wired, MIT Technology Review, Fortune, MSNBC, USA Today, Los Angeles Times, The Washington Post, Engadget, Gizmodo.
Speaker's Website
https://faculty.cc.gatech.edu/~dchau/
Host
Dominik Moritz