A HIGHLY EFFICIENT IMPLEMENTATION OF BACK-PROPAGATION ALGORITHM ON SIMD COMPUTERS

A version of the Back-Propagation (BP) algorithm is presented. It is fully realized with the matrix-vector product as basic operation (Level 2 BLAS), allowing portability onto various SIMD computers, with a degree of efficiency in the use of architecture which is equivalent to that characterizing matrix-vector implementation. An implementation is shown which reaches maximum efficiency on FPS M64 Series Minisupercomputers. The algorithm was tested on a FPS M64/30 (peak performances 12 MFLOPS) giving an asymptotic speed of 12 MFLOPS, corresponding to 6 MCPS (Million of Connections modified Per Second) in the feedforward step and to a maximum of 3 MCPS in the BP learning step. Significant applications using this algorithm are shown: linear separability, multiplexer and parity problem.