Parallel object-oriented software for training large-scale linear and nonlinear Support Vector Machines on multiprocessor systems. It implements a problem decomposition approach for the dual quadratic programming (QP) formulation, where the solver for the inner QP problems is based on parallel iterative gradient-projection methods. Both the data and the computations are distributed among the available processors. An effective caching strategy is used to save computational time in kernel evaluations. The MPI calls ensure portability to a wide range of strictly-coupled distributed-memory systems. The software can efficiently handle medium-to-large inner QP subproblems and allows to face very-large and even huge data set in reasonable time.
PGPDT - Parallel Gradient-Projection-based Decomposition Technique
ZANGHIRATI, Gaetano;
2005
Abstract
Parallel object-oriented software for training large-scale linear and nonlinear Support Vector Machines on multiprocessor systems. It implements a problem decomposition approach for the dual quadratic programming (QP) formulation, where the solver for the inner QP problems is based on parallel iterative gradient-projection methods. Both the data and the computations are distributed among the available processors. An effective caching strategy is used to save computational time in kernel evaluations. The MPI calls ensure portability to a wide range of strictly-coupled distributed-memory systems. The software can efficiently handle medium-to-large inner QP subproblems and allows to face very-large and even huge data set in reasonable time.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
