LOW-COST MASSIVELY PARALLEL NEUROCOMPUTING FOR PATTERN RECOGNITION IN MACROMOLECULAR SEQUENCES

CONNECTIONIST (NEURAL NETWORK) APPROACHES TO PATTERN RECOGNITION AND ANALYSIS HAVE ATTRACTED GREAT INTEREST RECENTLY BECAUSE OF THEIR FLEXIBILITY, ABILITY TO LEARN BY EXAMPLE, AND ABILITY TO "SELF-ORGANIZE" TO REVEAL HIDDEN PATTERNS AND RELATIONSHIPS IN THE INPUT DATA SET. NEURAL NETWORKS ARE INHERENTLY PARALLEL COMPUTATIONAL STRUCTURES AND THUS POTENTIALLY EXCELLENT CANDIDATES FOR IMPLEMENTATION ON MASSIVELY PARALLEL COMPUTERS. PARTICULARLY IN THE TRAINING STAGE, SERIAL IMPLEMENTATIONS OF CONNECTIONIST MODELS ARE OFTEN LIMITED WITH RESPECT TO EITHER THE SIZE OF A NETWORK THAT MAY PRACTICALLY BE SIMULATED OR THE NUMBER OF TRAINING TRIALS THAT ARE PRACTICAL. THE EXPLOSION IN THE SIZE OF MACROMOLECULAR SEQUENCE DATABASES (SUCH AS GENBANK) THAT IS ALREADY OCCURRING HAS CREATED A NEED FOR PATTERN ANALYSIS SOLUTIONS WITH COST-PERFORMANCE CHARACTERISTICS SUPERIOR TO THOSE CURRENTLY AVAILABLE. PHASE I OF THE PLANNED RESEARCH IMPLEMENTS EFFICIENT PARALLEL ALGORITHMS FOR ALL CRITICAL COMPONENTS OF A BASIC MULTILAYER, FEED-FORWARD NEURAL NETWORK MODEL. THESE INCLUDE A "LINEAR COMBINER" FOR THE MULTIPLICATION OF CONNECTION WEIGHT MATRICES WITH INPUT (OR ERROR) VECTORS, A SIGMOIDAL ACTIVATION FUNCTION TO INTRODUCE NONLINEARITY IN NEURON BEHAVIOR, AND A COMPLETE PARALLEL IMPLEMENTATION OF THE BACK-PROPAGATION (GENERALIZED DELTA RULE) TRAINING ALGORITHM. ONCE IMPLEMENTED, THE NETWORK IS BEING EVALUATED AGAINST ALTERNATIVE PARAMETERIZATIONS IN A PROTOTYPE DNA SEQUENCE PATTERN RECOGNITION TASK.