DEVELOPMENT OF A MULTI-DSP SYSTEM FOR NEURAL COMPUTATION

NEURAL COMPUTATION REQUIRES VERY HIGH COMPUTING BANDWIDTHS, MAINLY WHEN REAL-WORLD, TIME VARYING CLASSIFICATION PROBLEMS ARE ADDRESSED. THIS PROJECT OUTLINES A PLAN TO DEVELOP A LOW COST, MULTI-DIGITAL SIGNAL PROCESSING SYSTEM THAT INTERFACES TO UPSCALE PCS AND/OR WORKSTATIONS. THE MULTI-PROCESSOR ARCHITECTURE IS BASED ON A PARTITIONING SCHEME WHICH PRODUCES A COMPUTATION LOAD THAT SCALES WITH N(2), AND A GLOBAL COMMUNICATION LOAD (IDLE TIME) WHICH SCALES WITH RESPECT TO N (WHERE N IS THE NUMBER OF NEURONS IN A GIVEN LAYER). THE MULTI-PROCESSOR BOARD IS BEING TRANSPARENTLY INTEGRATED WITH AN EXISTING NEURAL NETWORK PACKAGE (NEUROSOLUTIONS) RECENTLY DEVELOPED AND COMMERCIALIZED, WHICH PROVIDES AN ICON-BASED SIMULATION ENVIRONMENT FOR DYNAMIC (AND STATIC) NEURAL NETWORKS. THE COMBINATION OF THE FAST HARDWARE WITH THE POWER OF NEUROSOLUTIONS PRODUCES AN EXCELLENT PRICE/PERFORMANCE RATIO WHICH SHOULD MAKE FAST NEURAL COMPUTING AFFORDABLE TO THE AVERAGE PC USER.