Automatic and Parallel Optimized Learning for Neural Networks performing MIMO Applications

Abstract

An automatic and optimized approach based on
\nmultivariate functions decomposition is presented to face
\nMulti-Input-Multi-Output (MIMO) applications by using
\nSingle-Input-Single-Output (SISO) feed-forward Neural
\nNetworks (NNs). Indeed, often the learning time and the
\ncomputational costs are too large for an effective use of MIMO
\nNNs. Since performing a MISO neural model by starting from
\na single MIMO NN is frequently adopted in literature, the
\nproposed method introduces three other steps: 1) a further
\ndecomposition; 2) a learning optimization; 3) a parallel
\ntraining to speed up the process. Starting from a MISO NN, a
\ncollection of SISO NNs can be obtained by means a multidimensional
\nSingle Value Decomposition (SVD). Then, a
\ngeneral approach for the learning optimization of SISO NNs is
\napplied. It is based on the observation that the performances of
\nSISO NNs improve in terms of generalization and robustness
\nagainst noise under suitable learning conditions. Thus, each
\nSISO NN is trained and optimized by using limited training
\ndata that allow a significant decrease of computational costs.
\nMoreover, a parallel architecture can be easily implemented.
\nConsequently, the presented approach allows to perform an
\nautomatic conversion of MIMO NN into a collection of
\nparallel-optimized SISO NNs. Experimental results will be
\nsuitably shown.