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:Approximate Equivariance SO(3) Needlet Convolution
- Kai Yi
DTSTART;TZID=Europe/London:20221110T170000
DTEND;TZID=Europe/London:20221110T180000
UID:TALK176786AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/176786
DESCRIPTION:This work focuses on the development of a rotation
-invariant needlet convolution for rotation group
SO(3) to distill multiscale information of spheric
al signals.\nThe spherical needlet transform is ge
neralized from $\\sS^2$ onto the SO(3) group\, whi
ch decomposes a spherical signal to approximate an
d detailed spectral coefficients by a set of tight
framelet operators. The spherical signal during t
he decomposition and reconstruction achieves rotat
ion invariance.\nBased on needlet transforms\, we
form a Needlet approximate Equivariance Spherical
CNN (NES) with multiple SO(3) needlet convolutiona
l layers. The network establishes a powerful tool
to extract geometric-invariant features of spheric
al signals.\nThe model allows sufficient network s
calability with multi-resolution representation. A
robust signal embedding is learned with wavelet s
hrinkage activation function\, which filters out r
edundant high-pass representation while maintainin
g approximate rotation invariance.\nThe NES achiev
es state-of-the-art performance for quantum chemis
try regression and Cosmic Microwave Background (CM
B) delensing reconstruction\, which shows great po
tential for solving scientific challenges with hig
h-resolution and multi-scale spherical signal repr
esentation.
LOCATION:Lecture Theatre 2
CONTACT:Pietro Lio
END:VEVENT
END:VCALENDAR