BEGIN:VCALENDAR METHOD:REQUEST PRODID:Microsoft Exchange Server 2010 VERSION:2.0 BEGIN:VTIMEZONE TZID:Eastern Standard Time BEGIN:STANDARD DTSTART:16010101T020000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;INTERVAL=1;BYDAY=1SU;BYMONTH=11 END:STANDARD BEGIN:DAYLIGHT DTSTART:16010101T020000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;INTERVAL=1;BYDAY=2SU;BYMONTH=3 END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT ORGANIZER;CN="Hunter, Tiffany":mailto:huntert1@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN=kzhao27@as u.edu:mailto:kzhao27@asu.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN='eecs_bscs @listserv.ohio.edu':mailto:eecs_bscs@listserv.ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN='eecs_bsee @listserv.ohio.edu':mailto:eecs_bsee@listserv.ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN='eecs_mscs @listserv.ohio.edu':mailto:eecs_mscs@listserv.ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN='eecs_phd@ listserv.ohio.edu':mailto:eecs_phd@listserv.ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN='eecs_msee @listserv.ohio.edu':mailto:eecs_msee@listserv.ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Abukamail, Nasseef":mailto:abukamai@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Allwine, D aniel":mailto:allwined@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Bartone, C hris":mailto:bartone@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Chenji, Ha rsha":mailto:chenji@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Goble, Jam es":mailto:goble@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Irwin, Den nis":mailto:irwind@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Jadwisienc zak, Wojciech":mailto:jadwisie@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Karanth, A vinash":mailto:karanth@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Kaya, Sava s":mailto:kaya@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Kelsey, Ra lph":mailto:kelsey@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Liu, Chang" :mailto:liuc@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Liu, Jundo ng":mailto:liuj1@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Mourning, Chad":mailto:mourning@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Ostermann, Shawn":mailto:osterman@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Rahman, Fa iz":mailto:rahmanf@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Schlicher, Jared":mailto:schliche@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Ugazio, Sa brina":mailto:ugazio@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Vasiliadis , Konstantinos":mailto:vassilia@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Welch, Lon nie":mailto:welch@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Zhu, Jim":m ailto:zhuj@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Wang, Zhew ei":mailto:wangz1@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Lindner, P atricia":mailto:lindnerp@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Ardrey, Gr egory":mailto:gardrey@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Yadav, Ani mesh":mailto:yadava@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Plis, Kevi n":mailto:plis@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Juedes, Da vid":mailto:juedes@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Fox, Patri ck":mailto:pfox@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Patterson, James":mailto:pattersj@ohio.edu ATTENDEE;ROLE=REQ-PARTICIPANT;PARTSTAT=NEEDS-ACTION;RSVP=TRUE;CN="Steinberg, Eric":mailto:steinber@ohio.edu DESCRIPTION;LANGUAGE=en-US:Title\nRevolutionizing Edge AI: Improved and Aut omated Model Compression\nBio\nKaiqi Zhao is a final-year Ph.D. candidate advised by Prof. Ming Zhao in the School of\nComputing and Augmented Intel ligence (SCAI) at Arizona State University (ASU). Her research\ninterests are machine learning (ML) and deep learning (DL) model compression as well as cloud\nand edge computing. Central to her research are innovating solu tions to optimize large-scale\nML/DL models for Internet-of-Things (IoT) d ata-driven applications. She has published\nfirst-authored papers at top-t ier AI conferences\, e.g.\, AISTATS (2023 and 2024)\, ICASSP (Oral)\,\nand InterSpeech (Oral)\, and top-tier edge computing conferences\, e.g.\, SEC \, and won the Best\nPoster Award at SEC’24. She did three research inte rnships at Amazon Web Services as an\nApplied Scientist. One of her intern ship works about compressing speech recognition models\nhas been integrate d into the Amazon Alex library for production usage. Additionally\, she wa s\nawarded the Graduate College Completion Fellowship\, the most prestigio us award for graduate\nstudents at ASU\, and SCAI Doctoral Fellowships in 2022 and 2023.\nAbstract\nMachine learning models are increasingly employe d by smart devices on the edge to support\nimportant applications such as real-time virtual assistants and privacy-preserving healthcare.\nHowever\, deploying state-of-the-art (SOTA) deep learning models on devices faces m ultiple\nserious challenges. First\, it is infeasible to deploy large mode ls on resource-constrained edge\ndevices whereas small models cannot achie ve the SOTA accuracy. Second\, it is difficult to\ncustomize the models ac cording to diverse application requirements in accuracy and speed and\ndiv erse capabilities of edge devices. This talk presents several novel soluti ons to\ncomprehensively address the above challenges through automated and improved model\ncompression. First\, it introduces Automatic Attention Pr uning (AAP)\, an adaptive\,\nattention-based pruning approach to automatic ally reduce model parameters while meeting\ndiverse user objectives in mod el size\, speed\, and accuracy. AAP achieves an impressive\n92.72% paramet er reduction in ResNet-101 on Tiny-ImageNet without causing any accuracy\n loss. Second\, it presents Self-Supervised Quantization-Aware Knowledge Di stillation\, a\nframework for reducing model precision without supervision from labeled training data. For\nexample\, it quantizes VGG-8 to 2 bits o n CIFAR-10 without any accuracy loss. Finally\, the talk\nexplores two mor e works\, Contrastive Knowledge Distillation and Module Replacing\, for fu rther\nimproving the performance of small models. All the works presented in this talk are designed to\naddress real-world challenges\, with applica tions extending to production environments such as\nAmazon Alexa\, and hav e been successfully deployed on diverse hardware platforms\, including\ncl oud instances and edge devices\, catalyzing AI for the edge.\n____________ ____________________________________________________________________\nMicr osoft Teams meeting\nJoin on your computer\, mobile app or room device\nCl ick here to join the meeting
View Site in Mobile | Classic
Share by: