BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.bham.ac.uk//v3//EN
BEGIN:VEVENT
CATEGORIES:Artificial Intelligence and Natural Computation se
minars
SUMMARY:Cost sensitive learning in high dimensions with ap
proximate nearest neighbours - Henry W J Reeve\, S
chool of Computer Science\, University of Manchest
er
DTSTART:20170206T110000Z
DTEND:20170206T120000Z
UID:TALK2374AT
URL:/talk/index/2374
DESCRIPTION:Speaker's homepage: http://www.cs.man.ac.uk/~reeve
h/\n\nHost: Dr Ata Kaban\n\nAbstract:\n\nWe shall
explore a nearest neighbour approach to multi-clas
s cost-sensitive learning in high dimensional feat
ure spaces. The behaviour of the classical k-neare
st neighbour algorithm has been explored extensive
ly\, culminating in the result of Chaudhuri and Da
sgupta which shows that the algorithm attains mini
max optimal learning rates in the binary classific
ation setting. However\, in high dimensional featu
re spaces exact nearest neighbour computations are
prohibitively expensive. Instead approximate near
est neighbours must be used. Kaban gave a bound on
the generalisation error of the approximate neare
st neighbour algorithm obtained by applying a 1-ne
arest neighbour classifier to randomly projected d
ata. The bound highlights the way in which random
projections enable the classifier to efficiently u
tilise low-dimensional structure within a high dim
ensional feature space. We consider approximate ne
arest neighbour algorithms in the context of the m
ore general problem of multi-class\, cost-sensitiv
e learning. We suppose that the data is concentrat
ed on a low-dimensional manifold within a high-dim
ensional ambient feature space and show that the c
ost-sensitive approximate nearest neighbour algori
thm is capable of exploiting this low dimensional
structure to attain minimax optimal learning rates
.
LOCATION:Computer Science\, The Sloman Lounge (UG)
CONTACT:Lars Kunze
END:VEVENT
END:VCALENDAR