Seminar Details
2015-10-20 (14.00) : SVD meets LS-SVM: a unifying picture
At Euler building (A002)
Organized by Mathematical Engineering
Speaker :
Johan Suykens (KU Leuven, ESAT)
Abstract :
In a first introductory part of this talk, the role of primal and dual model representations for data-driven modelling is explained.
Its relevance is illustrated e.g. with respect to sparsity, large scale learning algorithms, optimization modelling, and networks. Within this context
one may view least-squares support vector machines (LS-SVMs) as a class of core models in supervised and unsupervised learning, and beyond.
In a second part it is revealed how the Singular Value Decomposition (SVD) can be obtained within the LS-SVM setting. This is done by connecting to the Lanczos
decomposition theorem and the shifted eigenvalue problem. Also nonlinear SVD extensions are proposed. This presentation is based on the recent results of [Suykens, 2015].
Related reference:
Suykens J.A.K., ``SVD revisited: a new variational principle, compatible feature maps and nonlinear extensions'', Applied and Computational Harmonic Analysis, 2015, in press.