BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:1
UID:UW-Physics-Event-5425
DTSTART:20200506T160000Z
DTEND:20200506T170000Z
DTSTAMP:20260415T041153Z
LAST-MODIFIED:20200428T223512Z
LOCATION:Please register for this online event: http://physicsmeetsml.
 org
SUMMARY:Natural Graph Networks\, Physics ∩ ML Seminar\, Taco Cohen\,
  Qualcomm AI Research
DESCRIPTION:Message passing algorithms are the core of most neural net
 works that process information on graphs. Conventionally\, such method
 s are invariant under permutation of the messages and hence forget how
  the information flows through the network. Analyzing the local symmet
 ries of the graph\, we show that a more general message passing networ
 k can in fact be sensitive the flow of information by using different 
 kernels on different edges. This leads to an equivariant message passi
 ng algorithm that is more expressive than conventional invariant messa
 ge passing\, overcoming fundamental limitations of the latter. We deri
 ve the weight sharing and kernel constraints by modelling the symmetri
 es using elementary category theory and show that equivariant kernels 
 are “just” natural transformations between two functors. This gene
 ral formulation\, which we call Natural Networks\, gives a unified the
 ory to model many distinct forms of equivariant neural networks.
URL:https://www.physics.wisc.edu/events/?id=5425
END:VEVENT
END:VCALENDAR