BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.bham.ac.uk//v3//EN
BEGIN:VEVENT
CATEGORIES:Artificial Intelligence and Natural Computation se
minars
SUMMARY:A Simple and Consistent Technique for Vector-value
d Distribution Regression - Dr Zoltán Szabó\, Gats
by Computational Neuroscience Unit\, Centre for Co
mputational Statistics and Machine Learning (CSML)
University College London (UCL)
DTSTART:20150126T160000Z
DTEND:20150126T170000Z
UID:TALK1665AT
URL:/talk/index/1665
DESCRIPTION:I am going to tackle the regression task of vector
-valued outputs from probability distributions in
the two-stage sampled setting\, when only sets of
samples from the distributions are observable. The
studied distribution regression problem (DRP) cov
ers several important and challenging tasks in mac
hine learning and statistics\, including multi-ins
tance regression or point estimation problems (suc
h as hyperparameter identification). The inherent
two-stage sampled nature of the setup makes the de
rivation of theoretical performance guarantees rat
her difficult: to the best of our knowledge the on
ly available method from the large number of exist
ing techniques performs density estimation (which
typically performs poorly in practise)\, and restr
icts the problem to distributions with compact Euc
lidean support. In my talk\, I will present a simp
le\, ridge regression-based alternative to solving
the DRP problem: we embed the distribution to a r
eproducing kernel Hilbert space\, and learn the re
gressor from the embedded distribution to the outp
uts. We prove that under mild assumptions (on sepa
rable topological domains enriched with kernels)\,
this scheme is consistent\; moreover\, we derive
explicit rates of convergence in terms of the prob
lem difficulty. Specifically\, we prove that the s
et kernel is consistent in regression\, which was
a 15-year-old open\, and demonstrate the efficienc
y of our method in supervised entropy learning and
aerosol prediction based on multispectral satelli
te images. [Joint work with Bharath Sriperumbudur\
, Barnabas Poczos\, Arthur Gretton]\n\nPreprint: "
http://arxiv.org/abs/1411.2066"\n\nCode: "https://
bitbucket.org/szzoli/ite/"\n\nSpeaker's homepage:
http://www.gatsby.ucl.ac.uk/~szabo/
LOCATION:Mechanical Engineering\, B01
CONTACT:Lars Kunze
END:VEVENT
END:VCALENDAR