| + |
08/28/2026
12:45 - 1:45 PM |
Dr. Apurv Shukla
University of Michigan-Dearborn, USA |
Leaderboards with Receipts: Certifying LLM Rankings and Directing Preference Collection (in-person) |
|
|
Bio: Dr. Apurv Shukla is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Michigan–Dearborn. His research studies how autonomous decision-making systems operate under information, strategic, and physical constraints, with the goal of building safe, efficient, and economically aligned learning systems for critical infrastructure — including healthcare, power systems, markets, and platforms. His recent work develops the theory and algorithms of preference-based pure exploration: sequential experiment design for identifying best alternatives and Pareto frontiers from preference feedback. He got his PhD from Columbia University and was previously a postdoc at Texas A&M and University of Michigan, Ann Arbor.
Abstract: Every published model leaderboard invites two questions: is the ranking real, and how many more votes until it is? Current practice — Elo scores, uncalibrated confidence intervals, uniform vote collection — answers neither. In this talk I treat a leaderboard as what it is: a certificate in a fixed-confidence identification problem. Under a Bradley–Terry model, the information carried by a set of votes is a weighted graph Laplacian, and every ranking claim acquires a price: an Elo gap is certified at gap 2 over twice the effective resistance between the two models in the vote graph. This yields a procedure that certifies claims at confidence 1−δ , directs each next vote toward the current bottleneck, and stops when the claim set is certified. On Bradley–Terry ground truth fitted to 1.16M real Chatbot Arena battles, the certified best model costs about 10% of the votes required by confidence-interval practice, with zero errors across all runs — and the same machinery, pointed at preference annotation for alignment, certifies a helpful–harmless frontier at roughly 3× lower annotation cost than random collection.
|
| + |
09/25/2026
12:45 - 1:45 PM |
Dr. Tapadhir Das
University of the Pacific, USA |
Resilient AI against Adversarial Interactions in Autonomous Vehicular Perception (Remote) |
PDF |
|
Bio: Dr. Tapadhir Das is an Assistant Professor in the Department of Computer Science and Director of the Dependability And Security Laboratory at the University of the Pacific. His primary research focuses on protecting cyber-physical systems, such as intelligent vehicles and smart grid infrastructure, from cyberattacks and on developing resilient and trustworthy AI frameworks. As a researcher, Tapadhir has co-authored over 40 international conference papers, journal articles, and book chapters. He is the recipient of an IEEE CARS 2024 Best Paper Award and an IEEE IECON 2024 Best Paper Presentation Award. Tapadhir has also been recognized with the 2026 Hoefer Award for Outstanding Student-Faculty Research at the University of the Pacific, along with the 2026 Alumni Scientific Achievement Award from the Oregon Institute of Technology. He currently serves on the organizing and program committees of several IEEE conferences and is on the editorial and review boards for several high-impact journals.
Abstract: Autonomous Vehicles (AVs) are one of the most revolutionary developments in the field of vehicular technology, with companies like Waymo, Tesla, and Zoox leading the charge. These smart vehicles improve road safety, accessibility, and environmentally friendly driving practices through advanced sensing capabilities. However, the widespread adoption of this technology is currently hindered by major threats: adversarial attacks, intentional vandalism, and environmental degradation. These impact critical AV perception operations like pedestrian detection, sign recognition, and scene understanding. This presentation will discuss these current threats and how new & advanced AI methods can be leveraged to increase resiliency in AV perception systems against these evolving threats.
|
| + |
10/09/2026
12:45 - 1:45 PM |
Dr. Dilrukshi Gamage
University of Colombo School of Computing, Sri Lanka |
Trust and Truth in the Age of Generative AI: A Global South Perspective (Remote) |
PDF |
|
Bio: Dr. Dilrukshi Gamage is a Senior Lecturer at the University of Colombo School of Computing, Sri Lanka, where she leads the Colombo HCI Lab, and a Microsoft Research Fellow. Her research sits at the intersection of human–computer interaction and computational social science, focusing on the societal implications of AI-generated content — deepfakes, synthetic media, and misinformation and on building tools, labels, and literacy interventions that help people evaluate such content critically. She is a Visiting Research Fellow at the Centre for Protecting Women Online at The Open University, UK, and has held research affiliations with the Berkman Klein Center at Harvard and the Weizenbaum Institute in Berlin. Her work has been funded by competitive grants such as Google Academic Research Grant and MSR. She pioneered HCI as a research area in Sri Lanka and founded the Colombo ACM SIGCHI Chapter in 2019. Her work appears at ACM CHI, CSCW, ICWSM, IC2S2, Journal of Online Trust and Safety; she serves as a regular Associate Chair for CHI, Papers Chair for ACM COMPASS 2026, and on the editorial board of several key HCI journals.
Abstract: Generative AI has made it trivially easy to fabricate convincing images, audio, and yet the labels, detection tools, and literacy resources meant to help people navigate synthetic media are designed largely for Western platforms, languages, and users. In the Global South, where fact-checking infrastructure is thin and harms fall hardest on women and marginalized communities, this gap has serious consequences. This talk examines the problem through a series of empirical studies: user perceptions of warning-label designs for AI-generated content on social media; a field experiment in Sri Lanka evaluating Deepfake.help, a literacy intervention grounded in inoculation theory; and analyses of how online communities discuss, trade, and make sense of deepfakes. I will also consider how provenance standards such as C2PA and SynthID, and emerging regulation such as the EU AI Act, translate — or fail to translate into low-resource contexts. Drawing these threads together, I will argue for trust and safety infrastructure that centers Global South users, and outline pathways for building it through human-centered design, literacy, and policy.
|
| + |
10/16/2026
12:45 - 1:45 PM |
Dr. David Bianchi
National Center for Supercomputing Applications (NCSA), USA |
Requesting and Deploying NAIRR Allocations (Remote) |
PDF |
|
Bio: Dr. David Bianchi is a member of the NAIRR Pilot Allocations Working Group who leads the administration of the NAIRR Classroom Program. He is excited to help educators and researchers access resources to help achieve their educational goals and provide great learning experiences in AI education. Dr. David Bianchi is a research scientist at the National Center for Supercomputing Applications (NCSA) at the University of Illinois Urbana-Champaign. Prior to this, he completed his Ph.D. degree in the Department of Chemistry at UIUC, doing research in computational science, biophysical simulations, and metabolic engineering and synthetic biology. His work focuses on development of computational science platforms, analyses and research with applications ranging from chemistry to genomics, agriculture, health sciences, energy, synthetic biology and bioengineering, and applications of machine learning to these areas.
Abstract: The NAIRR Pilot aims to connect U.S. researchers and educators to the computational, data, and training resources needed to advance AI research and research that employs AI. In this seminar, we will discuss an overview of the NAIRR Program ecosystem, the various NAIRR project types and resources available, how to apply for a NAIRR allocation, frequently asked questions and common issues, and how to get started, get help, report on, and gain experience using NAIRR allocations once awarded.
|
| + |
10/23/2026
12:45 - 1:45 PM |
Dr. Traci Carte
Illinois State University, USA |
Encouraging Agentic AI Use in Higher Education: A Repeated-Measures Study of Student Attitudes and Intentions (in-person) |
PDF |
|
Bio: Dr. Traci Carte earned her PhD degree from University of Georgia and she is curently a professor of information systems at School of Information Technology, Illinois State University, USA. Recently, she concluded a six-year term as director of the School of Information Technology and previously served as department chair at Kennesaw State University. During her time in administrative roles, she successfully led ABET reaccreditation twice and spent considerable time on curriculum development. She has been recognized by her peers and former students with 15 separate teaching awards. Her research expertise is in the area of collaboration using technology, and her work has been published in such journals as MIS Quarterly, Information Systems Research, and the Journal of AIS. She received an MISQ best paper award, an AMCIS best paper award and multiple outstanding reviewer awards. In 2024 she was named an AIS Fellow. She provides leadership to the discipline through the Association of Information Systems where she has served as a program chair for both ICIS and AMCIS as well as serving as a track chair at ICIS multiple times, a mini track chair at AMCIS multiple times, and as a frequent reviewer. Most recently, she served 4 years as the VP of Conferences for AIS.
Abstract: Artificial intelligence (AI) tools are increasingly embedded in higher education possibly leading to enhanced learning outcomes when used to support, rather than replace, student cognition. However, inappropriate use may undermine critical thinking and promote over-reliance. To better understand how students engage with AI and to encourage more effective use, this study examines whether a structured intervention can influence student attitudes and intentions toward AI. Drawing on the Technology Acceptance Model (TAM) and Expectation-Confirmation Theory, we conducted a repeated-measures investigation across two contexts: a low-stakes laboratory experiment and a higher-stakes field study. The intervention emphasized guided prompt engineering strategies designed to position AI as a learning support tool rather than a shortcut to solutions. In low-stakes conditions with limited task knowledge, AI use reinforced positive attitudes, consistent with expectation confirmation. In contrast, in higher-stakes conditions where students had greater task familiarity, AI use led to negative disconfirmation and declining attitudes. Across both settings, behavioral intentions remained relatively stable, suggesting greater resistance to short-term change. These findings highlight the importance of modeling purposefully, agentic AI use in educational settings.
|
| + |
11/06/2026
12:45 - 1:45 PM |
Arman Behnam
Illinois Institute of Technology, USA |
Asking Machines Why: Explanation and Reasoning in Graphs, Memory, and Language Models (in-person) |
PDF |
|
Bio: Arman Behnam is a Ph.D. candidate in Computer Science at Illinois Institute of Technology, advised by Dr. Binghui Wang. His research is in machine learning, and concerns making explanation and reasoning in AI systems something that can be checked. His work has appeared at NeurIPS 2026, NeurIPS 2025, ECCV 2024, and IEEE BIBM 2024, and he serves as a reviewer for NeurIPS, ICML, ICLR, and AAAI. He has been an AI research scientist intern at Mayo Clinic's Department of Artificial Intelligence and Informatics, and is a co-founder of Quis, Inc., where he works on memory systems for AI agents. He will defend his dissertation in November 2026.
Abstract: When a machine learning system gives us an answer, we usually ask whether it is right. Increasingly we want to ask something harder: why did this model flag this molecule? Which stored fact did the assistant actually use to answer me? Did the model reason, or did it recall? This talk follows that question through three settings, each built from the ground up with no background assumed. We begin with graphs and graph neural networks, what they are and what it means to explain a prediction made over a structure where everything connects to everything else. We turn next to memory and retrieval, how modern systems store what they learn, how they search it, and how you would ever know which stored item actually helped. Finally, language models, what they are and what it would mean for one to reason and explain.
|
| + |
11/20/2026
12:45 - 1:45 PM |
Dr. Angelo Geninatti Cossatin
University of Turin, Italy |
Sustainable Recommender Systems: From User Interface to Algorithm (Remote) |
PDF |
|
Bio: Dr. Angelo Geninatti Cossatin is a Postdoctoral Researcher at the Computer Science Department of the University of Turin where he obtained a PhD in Computer Science with Honors. His research mainly focuses on sustainability and social good for intelligent systems, spanning Human-Centered AI, user control, and reducing the computational resources required by recommender systems. He has been a visiting researcher at the University of Copenhagen, where he worked on reproducibility in information retrieval and on fairness in recommender systems. He is the author of several research articles related to intelligent systems and recommender systems, and a program committee member of the top conferences in his research areas. He has been co-chair of the second edition of the workshop on Information Retrieval for Underserved Users (IR4U2), and of the RecSoGood workshop.
Abstract: We examine both user-facing and system-level sustainability improvements in recommendation technology. The first part investigates how digital nudges and interface design can influence consumers toward more sustainable fashion choices, with experimental findings showing significantly increased conversion rates to second-hand items through visual sustainability labels. The second part introduces a novel Multimodal Dimensionality Reduction (MDR) technique that condenses high-dimensional data without compromising on recommendation quality, resulting in substantial computational resource savings across multiple datasets and algorithms. The goal is to make recommender systems more "sustainable by design" at both the user experience and algorithmic levels.
|
