BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:iCalendar-Ruby
BEGIN:VEVENT
CATEGORIES:Conferences/Seminars/Lectures
DESCRIPTION:Title: Feature Geometry guides Model Design in Deep Learning\n\
 nAbstract: The geometry of learned features can provide crucial insights on
  model design in deep learning. In this talk\, we discuss two recent lines 
 of work that reveal how the evolution of learned feature geometry during tr
 aining both informs and is informed by architecture choices. First\, we exp
 lore how deep neural networks transform the input data manifold by tracking
  its evolving geometry through discrete approximations via geometric graphs
  that encode local similarity structure. Analyzing the graphs’ geometry rev
 eals that as networks train\, the models’ nonlinearities drive geometric tr
 ansformations akin to a discrete Ricci flow. This perspective yields practi
 cal insights for early stopping and network depth selection informed by dat
 a geometry. The second line of work concerns learning under symmetry\, incl
 uding permutation symmetry in graphs or translation symmetry in images. Gro
 up-convolutional architectures can encode such structure as inductive biase
 s\, which can enhance model efficiency. However\, with increased depth\, co
 nventional group convolutions can suffer from instabilities that manifest a
 s loss of feature diversity. A notable example is oversmoothing in graph ne
 ural networks. We discuss unitary group convolutions\, which provably stabi
 lize feature evolution across layers\, enabling the construction of deeper 
 networks that are stable during training.
DTEND:20260129T151500Z
DTSTAMP:20260307T112529Z
DTSTART:20260129T141500Z
GEO:42.361613;-71.092293
LOCATION:Building 45 (MIT Stephen A. Schwarzman College of Computing)\, 230
SEQUENCE:0
SUMMARY:Plenary Talk | Prof. Melanie Weber
UID:tag:localist.com\,2008:EventInstance_51844438758910
URL:https://calendar.mit.edu/event/plenary-talk-prof-melanie-weber
END:VEVENT
END:VCALENDAR