BEGIN:VCALENDAR VERSION:2.0 PRODID:icalendar-ruby CALSCALE:GREGORIAN X-WR-CALNAME:MS Non-thesis (Project) Final Exam - David Smerkous X-WR-TIMEZONE:Pacific Time (US & Canada) BEGIN:VEVENT DTSTAMP:20260605T224104Z UID:tag:localist.com\,2008:EventInstance_48935192129017 DTSTART:20250224T180000Z DTEND:20250224T190000Z DESCRIPTION:TITLE: Enhancing Diversity in Bayesian Deep Learning via Hypers pherical Energy Minimization of CKA\n\nABSTRACT: Particle-based Bayesian d eep learning often requires a similarity metric to compare two networks. H owever\, naive similarity metrics lack permutation invariance and are inap propriate for comparing networks. Centered Kernel Alignment (CKA) on featu re kernels has been proposed to compare deep networks but has not been use d as an optimization objective in Bayesian deep learning. In this paper\, we explore the use of CKA in Bayesian deep learning to generate diverse en sembles and hypernetworks that output a network posterior. Noting that CKA projects kernels onto a unit hypersphere and that directly optimizing the CKA objective leads to diminishing gradients when two networks are very s imilar. We propose adopting the approach of hyperspherical energy (HE) on top of CKA kernels to address this drawback and improve training stability . Additionally\, by leveraging CKA-based feature kernels\, we derive featu re repulsive terms applied to synthetically generated outlier examples. Ex periments on both diverse ensembles and hypernetworks show that our approa ch significantly outperforms baselines in terms of uncertainty quantificat ion in both synthetic and realistic outlier detection tasks.\n\nMAJOR ADVI SOR: Fuxin Li\nCOMMITTEE: Xiao Fu\nCOMMITTEE: Alan Fern\nCOMMITTEE: Stefan Lee GEO:44.567164;-123.278692 LOCATION:Kelley Engineering Center\, 1005 SUMMARY:MS Non-thesis (Project) Final Exam - David Smerkous URL;VALUE=URI:https://events.oregonstate.edu/event/ms-non-thesis-project-fi nal-exam-david-smerkous CATEGORIES:Lecture or Presentation END:VEVENT END:VCALENDAR