Learning Neural Network Circuit based on Logarithmic Multipliers

Abstract

Models for artificial intelligence, machine learning, and neural networks are implemented on digital computers with a von Neumann architecture. Few studies have considered analog neural networks. Our previous study used multipliers to represent connecting weights in a neural network. The multipliers calculate the product of input signals and their corresponding connecting weights. However, using MOSFET multipliers, their operating range is limited by semiconductor characteristics. The input and output ranges for networks that use these multipliers are thus limited. Furthermore, the circuit operation becomes unstable. Here, we propose a logarithmic four-quadrant multiplier for representing connecting weights. The output of this multiple circuit is a more accurate value compared to the previous circuit. Experiments show that this multiplier exhibits stable operation over a wide range. Therefore, this model can be used directly for input/output of an analog control unit. A model that uses only analog electronic circuits is presented. Its learning time is quite short compared to that for models implemented on a digital computer. We increased the number of units and network layers. We suggest the possibility of a hardware implementation of a deep learning model. Furthermore, this model expects the elucidation of the biomedical learning mechanism.