BEGIN:VCALENDAR VERSION:2.0 X-WR-CALNAME:EventsCalendar PRODID:-//hacksw/handcal//NONSGML v1.0//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York LAST-MODIFIED:20240422T053451Z TZURL:https://www.tzurl.org/zoneinfo-outlook/America/New_York X-LIC-LOCATION:America/New_York BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT CATEGORIES:College of Engineering DESCRIPTION:Topic: “Generative AI for Explainable, Robust, and Trustworth y Edge Intelligence” Abstract: The proliferation of edge devices has cre ated a vast, distributed, and privacy-sensitive data ecosystem—the new f rontier for AI. However, this opportunity intersects with a fundamental co nstraint: the AI community is entering a "peak-data" era, where the supply of high-quality web data is becoming exhausted. Major AI labs including G oogle, OpenAI, Anthropic, and Meta are already integrating synthetic data into their training pipelines, but translating this capability to real-wor ld edge environments introduces significant system-level challenges. This talk presents a new paradigm for Trustworthy Edge Intelligence, using Gene rative AI and Robust Learning as core primitives to create efficient, trus tworthy, and scalable AI systems. Dr. Vahidian’s research addresses thre e critical challenges through integrated solutions: engineering generative AI to grow synthetic corpus coverage without exploding GPU, memory, or co mmunication usage on constrained edge intelligence systems; using robust l earning to diagnose and correct failures on underrepresented subgroups of data; and developing protocols that enable collaboration without raw data movement where data is siloed at the edge. These solutions bring together efficient and explainable generative AI for synthetic data pipelines, robu st learning that targets domain shift and model failures on underrepresent ed data subgroups in edge intelligence systems where data lives—using fe derated learning, where I introduced the first exploration of federated in struction tuning (FedIT) for LLMs , with resource-efficient protocols for heterogeneous devices. Conducted within the NSF AI Institute for Edge Comp uting (Athena), our works integrate hardware-aware safeguards, advancing a future where edge devices collaboratively create and refine models using generative AI—ensuring robust performance without compromising privacy. Biography: Dr. Saeed Vahidian conducted his Postdoctoral research at Duke University, with Prof. Yiran Chen, Director of the NSF AI Institute for Ed ge Computing (Athena) —one of the 27 National AI Institutes established by the U.S. National Science Foundation with $20,000,000 in federal fundin g. He received his Ph.D. in Electrical and Computer Engineering from the U niversity of California San Diego (UCSD). His research sits at the interse ction of Generative AI and Efficient Edge Intelligence, developing hardwar e-aware algorithms, robust learning methods, and synthetic data generation for vision-language, multimodal, and video pipelines. He has collaborated with Qualcomm AI and academic institutions across the U.S., Canada, and E urope—efforts that led to an invitation from NASA to contribute to a pro ject on Edge Intelligence. His publications appear in NeurIPS, ICLR, CVPR, ECCV, ICCV, UAI, IEEE Transactions on AI, and JMLR. He has served as Chai r at CVPR workshops and as a reviewer for ICML, NeurIPS, CVPR, etc. The Re search Presentation is open to the public free of charge. *For further inf ormation, \nEvent page: https://www.umassd.edu/events/cms/speaker-dr-saeed -vahidian-postdoctoral-scholar-duke-university.php X-ALT-DESC;FMTTYPE=text/html:

Topic: “Gene rative AI for Explainable\, Robust\, and Trustworthy Edge Intelligence”< /p>\n

Abstract: The proliferation of edge devices has c reated a vast\, distributed\, and privacy-sensitive data ecosystem—the n ew frontier for AI. However\, this opportunity intersects with a fundament al constraint: the AI community is entering a "peak-data" era\, where the supply of high-quality web data is becoming exhausted. Major AI labs inclu ding Google\, OpenAI\, Anthropic\, and Meta are already integrating synthe tic data into their training pipelines\, but translating this capability t o real-world edge environments introduces significant system-level challen ges. This talk presents a new paradigm for Trustworthy Edge Intelligence\, using Generative AI and Robust Learning as core primitives to create effi cient\, trustworthy\, and scalable AI systems.

Dr. Vahidian’s re search addresses three critical challenges through integrated solutions: e ngineering generative AI to grow synthetic corpus coverage without explodi ng GPU\, memory\, or communication usage on constrained edge intelligence systems\; using robust learning to diagnose and correct failures on underr epresented subgroups of data\; and developing protocols that enable collab oration without raw data movement where data is siloed at the edge. These solutions bring together efficient and explainable generative AI for synth etic data pipelines\, robust learning that targets domain shift and model failures on underrepresented data subgroups in edge intelligence systems w here data lives—using federated learning\, where I introduced the first exploration of federated instruction tuning (FedIT) for LLMs \, with resou rce-efficient protocols for heterogeneous devices. Conducted within the NS F AI Institute for Edge Computing (Athena)\, our works integrate hardware- aware safeguards\, advancing a future where edge devices collaboratively c reate and refine models using generative AI—ensuring robust performance without compromising privacy.

Biography: Dr. Saee d Vahidian conducted his Postdoctoral research at Duke University\, with P rof. Yiran Chen\, Director of the NSF AI Institute for Edge Computing (Ath ena) —one of the 27 National AI Institutes established by the U.S. Natio nal Science Foundation with $20\,000\,000 in federal funding. He received his Ph.D. in Electrical and Computer Engineering from the University of Ca lifornia San Diego (UCSD). His research sits at the intersection of Genera tive AI and Efficient Edge Intelligence\, developing hardware-aware algori thms\, robust learning methods\, and synthetic data generation for vision- language\, multimodal\, and video pipelines. He has collaborated with Qual comm AI and academic institutions across the U.S.\, Canada\, and Europe— efforts that led to an invitation from NASA to contribute to a project on Edge Intelligence. His publications appear in NeurIPS\, ICLR\, CVPR\, ECCV \, ICCV\, UAI\, IEEE Transactions on AI\, and JMLR. He has served as Chair at CVPR workshops and as a reviewer for ICML\, NeurIPS\, CVPR\, etc.

\ n

The Research Presentation is open to the public free of charge.

\n< p>*For further information\,

Event page: https://www.umassd.edu/events/cms/speaker-dr-saeed-vahidian-pos tdoctoral-scholar-duke-university.php

DTSTAMP:20260429T201552 DTSTART;TZID=America/New_York:20260323T111500 DTEND;TZID=America/New_York:20260323T121500 LOCATION:Lester W. Cory Conference Room, Science & Engineering Building (SENG), Room 213A SUMMARY;LANGUAGE=en-us:Speaker: Dr. Saeed Vahidian, Postdoctoral Scholar, D uke University UID:15536bdb2ee4f5b6ff7b1e93011407bc@www.umassd.edu END:VEVENT END:VCALENDAR