Use of Level 3 BLAS kernels in neural networks: the Back-Propagation algorithm

The Back-Propagation (BP) algorithm is analyzed from a computational point of view, showing the advantages of the training in batching mode. This approach makes possible the use of Level 3 BLAS, allowing portability with high performance onto SIMD and MIMD computers. The proposed implementation is tested on FPS M64 Series Minisupercomputers where the peak performances of the processor are reached. The asymptotic speed in the learning mode is 38 MFLOPS or 9.5 MCPS (for FPS M64/60). Examples of applications of the algorithm are shown.