Nesta sexta-feira 19/09/2003 as 11hs, teremos a presenca do Prof. von Zuben que apresentara o seminario:
Titulo: Machine Learning in practice: Support Vector Machines and Ensembles
Resumo:
Support vector machines (SVMs) comprehend a relatively new type of
supervised learning algorithm, originally introduced by Vapnik and
successively extended by a number of researchers. Their remarkably
robust performance with respect to sparse and noisy data is making
them
the system of choice when one has to deal with nontrivial
classification, regression, and even forecasting problems. SVMs tackle
these problems by (non)linearly mapping input data into high-dimensional
feature spaces, wherein a linear decision surface is designed. SVMs
comprise an approximate implementation of the structural risk
minimization, which asserts that the generalization error is delimited
by the sum of the training error and a parcel that depends on the
Vapnik-Chervonenkis (VC) classifier dimension. By minimizing this sum,
high generalization performance may be obtained. Another important
feature of the support vector learning approach is that the underlying
optimization problems are inherently convex and have no local minima,
which comes as the result of applying Mercer's conditions on the
characterization of kernels. Furthermore, the number of free parameters
in an SVM does not explicitly depend upon the input dimensionality
of
the problem at hand, something very interesting when untangling complex
input data relationships.
By other means, ensembles involve the generation, selection, and
linear/nonlinear combination of a set of individual tools designed
to
simultaneously cope with the same task. This is typically done through
the variation of some configuration parameters and/or employment of
different training procedures. Such ensembles, a.k.a. committees, should
properly integrate the knowledge embedded in the component tools
(devised to provide redundancy), and have frequently produced more
accurate and robust models.
Some proposals to synthesize extended SVM models will be described,
and the performance when dealing with prediction, regression and
classification problems will be presented and discussed.
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | Ronaldo Dias | | Universidade Estadual de Campinas, UNICAMP | | Departamento de Estatistica, IMECC | | email address: dias@ime.unicamp.br | | URL: http://www.ime.unicamp.br/~dias | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~