BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Artificial Intelligence Research Group Talks (Comp
uter Laboratory)
SUMMARY:Graph Neural Networks for Geometric Graphs - Chait
anya K. Joshi\, Simon V. Mathis
DTSTART;TZID=Europe/London:20221108T130000
DTEND;TZID=Europe/London:20221108T140000
UID:TALK183641AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/183641
DESCRIPTION:Join us in Lecture Theatre 2 or on "Zoom":https://
zoom.us/j/99166955895?pwd=SzI0M3pMVEkvNmw3Q0dqNDVR
alZvdz09\n\nGeometric graphs are spatially embedde
d graphs used to model systems in biochemistry\, p
hysical simulations and multiagent robotics. Impor
tantly\, graph attributes transform along with glo
bal Euclidean transformations or symmetries of the
system\, such as rotations\, reflections\, and tr
anslation. Graph Neural Networks (GNNs) with globa
l symmetries 'baked in' have emerged as the archit
ecture of choice for geometric graphs. This talk w
ill introduce two classes of Geometric GNNs: (1) *
Equivariant GNNs*\, which use both scalar and geom
etric features that are equivariant to global symm
etries\; and (2) *Invariant GNNs*\, which only rea
son locally via invariant scalars such as distance
s and angles. Additionally\, we will study the exp
ressive power of the two classes of Geometric GNNs
from the perspective of distinguishing geometric
graphs\, i.e. graph isomorphism testing. We will i
ntroduce a *Geometric Weisfeiler-Leman* graph isom
orphism test (GWL). We will then use the GWL frame
work to formally show that equivariant GNNs have g
reater expressive power than invariant GNNs\, as t
hey enable propagating geometric information beyon
d local neighbourhoods and compositionally build l
ong-range interactions.\n\nThis talk is based on t
he paper *"On the Expressive Power of Geometric Gr
aph Neural Networks"*\, by Chaitanya K. Joshi (x)\
, Cristian Bodnar (x)\, Simon V. Mathis\, Taco Coh
en\, and Pietro LiĆ²\, to be presented as an Oral a
t the _NeurIPS 2022 Workshop on Symmetry and Geome
try in Neural Representations_.\n\n\n\n
LOCATION:Lecture Theatre 2 and Zoom
CONTACT:Mateja Jamnik
END:VEVENT
END:VCALENDAR