BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260525T171231EDT-8550z4e5H1@132.216.98.100
DTSTAMP:20260525T211231Z
DESCRIPTION:Abstract\n\nLearning on graphs has emerged as a fundamental dir
 ection in machine learning\, as part of the broader field of geometric dee
 p learning\, which focuses on modeling the intricate structures and symmet
 ries underlying complex data.\n	However\, existing standard paradigm of gra
 ph neural networks (GNNs) -- message passing neural networks (MPNNs) -- fa
 ce inherent limitations in terms of expressivity\, capacity\, flexibility\
 , and generality.\n	This thesis introduces a novel paradigm—learning on pse
 udo-coordinates—which provides not only enhanced expressive and representa
 tional power but also greater modeling flexibility.\n	Pseudo-coordinates se
 rve as coordinate-like spatial representations that relax the strict const
 raints of conventional coordinate systems.\n	They are particularly valuable
  for domains lacking canonical coordinate systems\, such as graphs and man
 ifolds.\n	The main body of this thesis is composed of three principal studi
 es.\n	The first study focuses on the design of graph Transformers\, represe
 nting a realization of the graph learning paradigm on pseudo-coordinates.
 \n	In this work\, we introduce a powerful graph Transformer\, encompassing 
 a novel pseudo-coordinate design and an advanced model architecture\, whic
 h demonstrates the superior effectiveness of graph learning on pseudo-coor
 dinates compared to the conventional message-passing paradigm.\n	The second
  study investigates alternative approaches within this paradigm.\n	We intro
 duce a novel graph convolutional operator defined on pseudo-coordinates\, 
 leveraging the design of continuous convolutional kernels.\n	The proposed g
 raph convolution is not only highly expressive but also generalizes many e
 xisting GNN formulations\, including but not limited to MPNNs and polynomi
 al spectral GNNs.\n	More importantly\, it exhibits distinct characteristics
  from attention mechanisms in graph Transformers\, thereby expanding the d
 esign space for developing diverse and powerful graph neural architectures
 .\n	The final study\, inspired by recent advances in Transformer-based mult
 imodal foundation models\, revisits the design of graph Transformers. We d
 emonstrate that\, contrary to the prevailing trend of introducing complex 
 architectural modifications\, plain Transformer architectures—when equippe
 d with our proposed lightweight enhancements—can serve as highly effective
  graph learners. These enhancements are broadly applicable across domains 
 and gracefully reduce to the original Transformer when necessary. This fin
 ding underscores the versatility of vanilla Transformer architectures and 
 highlights their strong potential as a unified backbone for multimodal lea
 rning across language\, vision\, and graph domains.\n	Collectively\, this t
 hesis underscores the broad potential of the graph learning on pseudo-coor
 dinates paradigm\, which offers not only strong expressivity and flexibili
 ty but also a promising pathway toward unifying graph learning with multim
 odal foundation models.\n
DTSTART:20260114T193000Z
DTEND:20260114T213000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Liheng Ma – Towards More Expressive Graph Neural Net
 works via Learning on Pseudo-Coordinates
URL:/ece/channels/event/phd-defence-liheng-ma-towards-
 more-expressive-graph-neural-networks-learning-pseudo-coordinates-370197
END:VEVENT
END:VCALENDAR