A memory-based artificial neural network

The authors propose a novel memory-based architecture of artificial neural networks, namely, the method of multiplexing the memory table for approximate realization of the sigmoidal function. The lack of precision is compensated by the host computer (which serves as the development tool) when the system works in the learning stage. Moreover, the weight of a neuron can be changed smoothly since the memory is modified to be writable in broadcasting for a selected address bit. As a result, the learning can be made fast with a relatively low cost. The execution time for the real input is very fast, especially when the input is binary, because the neural operation is achieved only by once table-looking. Implementation of the proposed architecture is presented