A magnetic neural network utilizing universal arithmetic modules for pulse-train signal processing

The authors consider two types of magnetic neural network architecture based on pulse-train signal processing with high reliability. One is realized by neuron units and synapse units, both of which use only a universal arithmetic module (UAM) having adder, multiplier, and delay (memory) functions all in one unit. This architecture thus results in an extremely homogeneous implementation. The other is realized by dividing the function of the UAM into a neuron unit and a synapse unit. The neuron unit provides functions of nonlinear adder and memory by using the properties of the magnetic core such as pulse storage, integration, and nonlinearity