SIMPLIFIED QUASI-NEWTON PROJECTION METHOD ARITHMETIC SYSTEM, NEURAL NETWORK LEARNING SYSTEM, RECORDING MEDIUM, AND SIGNAL PROCESSOR

摘要

PROBLEM TO BE SOLVED: To provide the system capable of shortening calculation time, reducing the consumption of a memory and accurately obtaining a calculation result at the time of the learning of a neural network by applying a learning quasi- Newton projection method to the neural network which has M synapses (combination), sets an upper-limit and a lower-limit value for each synapse load, and performs learning so that specific conditions are met. ;SOLUTION: This neural network learning system is provided with the neural network 12, a learning control part 14, and a standard pattern storage part 18. Here, the neural network 12 uses a rewritable memory such as a RAM and an EEPROM. Further, the learning control part 14 is constituted as a computer device and a CPU as its kernel uses a digital arithmetic unit. Then the learning quasi-Newton projection method is applied to the neural network which has M synapses, sets upper-limit values B_iU and lower-limit values B_iL for the respective synapse loads W_i (i=1, 2...M), and performs the learning so that B_iL≤W_i≤B_iU holds.;COPYRIGHT: (C)1999,JPO