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:Laboratory for Scientific Computing
SUMMARY:Kernels for Sequentially Ordered Data - Dr Franz K
iraly
DTSTART;TZID=Europe/London:20161125T110000
DTEND;TZID=Europe/London:20161125T120000
UID:TALK69355AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/69355
DESCRIPTION:Kernel learning is a general framework providing m
ethodology for descriptive/exploratory statistics\
, non-linear regression and classification learnin
g\, for objects of any kind\, for example time ser
ies\, text\, matrices\, vectors up to mirror symme
tries\, and so on - given the right feature repres
entation encoded by the so-called kernel\, a non-l
inear scalar product. After briefly reviewing the
kernel learning framework and prior work on learni
ng with structured objects or invariances\, we pre
sent methodological foundations for dealing with s
equential data of any kind\, such as time series\,
sequences of graphs\, or strings. Our approach i
s based on signature features which can be seen as
an ordered variant of sample (cross-)moments\; it
allows to obtain a "sequentialized" version of an
y static kernel. The sequential kernels are effic
iently computable for discrete sequences and are s
hown to approximate a continuous moment form in a
sampling sense. A number of known kernels for sequ
ences arise as "sequentializations" of suitable st
atic kernels: string kernels may be obtained as a
special case\, and alignment kernels are closely
related up to a modification that resolves their
open non-definiteness issue. Our experiments indi
cate that our signature-based sequential kernel fr
amework may be a promising approach to learning wi
th sequential data\, such as time series\, that al
lows to avoid extensive manual pre-processing.
LOCATION:Rayleigh seminar room\, 2nd floor\, Maxwell Buildi
ng\, Cavendish laboratory
CONTACT:Martine Gregory-Jones
END:VEVENT
END:VCALENDAR